!gdZddlmZmZddlmZddlmZddlZ ddl Z ddl m Z mZmZddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZddlmZddl m!Z!ddl"m#Z#ddlm$Z$ddl%m&Z&ddl'm(Z(m)Z)ddl*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0ddl1m2Z2m3Z3m4Z4ddl5m6Z6ddl7m8Z8ddl9m:Z:ddl;mGdde=Z?Gdde(jZAGd d!e(jZCy)"z; Dynamic factor model. Author: Chad Fulton License: BSD-3 ) MONTH_END QUARTER_END) OrderedDict)warnN) cho_factor cho_solve LinAlgError)_is_using_pandas)int_like)cache_readonly)OLS)GLM)PCA)SARIMAX) QuarterlyAR1)VAR)Bunch) string_like)lagmat)mlemodelinitialization)companion_matrix is_invertibleconstrain_stationary_univariate!constrain_stationary_multivariate!unconstrain_stationary_univariate#unconstrain_stationary_multivariate)SMOOTHER_STATESMOOTHER_STATE_COVSMOOTHER_STATE_AUTOCOV) PandasData) SimpleTable)Summary) fmt_paramscfeZdZdZdZedZedZedZedZ edZ y) FactorBlocka Helper class for describing and indexing a block of factors. Parameters ---------- factor_names : tuple of str Tuple of factor names in the block (in the order that they will appear in the state vector). factor_order : int Order of the vector autoregression governing the factor block dynamics. endog_factor_map : pd.DataFrame Mapping from endog variable names to factor names. state_offset : int Offset of this factor block in the state vector. has_endog_Q : bool Flag if the model contains quarterly data. Notes ----- The goal of this class is, in particular, to make it easier to retrieve indexes of subsets of the state vector that are associated with a particular block of factors. - `factors_ix` is a matrix of indices, with rows corresponding to factors in the block and columns corresponding to lags - `factors` is vec(factors_ix) (i.e. it stacks columns, so that it is `factors_ix.ravel(order='F')`). Thinking about a VAR system, the first k*p elements correspond to the equation for the first variable. The next k*p elements correspond to the equation for the second variable, and so on. It contains all of the lags in the state vector, which is max(5, p) - `factors_ar` is the subset of `factors` that have nonzero coefficients, so it contains lags up to p. - `factors_L1` only contains the first lag of the factors - `factors_L1_5` contains the first - fifth lags of the factors c||_t|j|_||_|jdd|f|_||_||_|jdkDrtd|j|_ n|j|_ |j|jz|_ |j|d<|j|d<|j|d<|j|d<|j|d<y)Nrfactors factors_ar factors_ix factors_L1 factors_L1_5) factor_nameslen k_factors factor_orderlocendog_factor_map state_offset k_endog_Qmax _factor_orderk_statesr)r*r+r,r-)selfr.r1r3r4r5s i/var/www/dash_apps/app1/venv/lib/python3.12/site-packages/statsmodels/tsa/statespace/dynamic_factor_mq.py__init__zFactorBlock.__init__Qs(T../( 0 4 4Q _ E(" >>A !$Q(9(9!:D !%!2!2D ););; ,,Y!__\!__\!__\#00^c|j}tj|tj|j|j zz|j |jfj S)z3Factor state index array, shaped (k_factors, lags).)r4npreshapearanger0r7Tr9os r:r+zFactorBlock.factors_ixfsX   zz!bii9K9K(KLL--t~~>@@A Br<ct|j}tj|||j|jzzS)z5Factors and all lags in the state vector (max(5, p)).)r4r>s_r0r7rBs r:r)zFactorBlock.factorsrs4   uuQq4>>D,>,>>>??r<ct|j}tj|||j|jzzS)z;Factors and all lags used in the factor autoregression (p).)r4r>rEr0r1rBs r:r*zFactorBlock.factors_arys4   uuQq4>>D,=,===>>r<cZ|j}tj|||jzS)z!Factors (first block / lag only).r4r>rEr0rBs r:r,zFactorBlock.factors_L1s)   uuQq4>>)**r<c`|j}tj|||jdzzS)zFactors plus four lags.r(rHrBs r:r-zFactorBlock.factors_L1_5s.   uuQq4>>A--..r<N) __name__ __module__ __qualname____doc__r;propertyr+r)r*r,r-r<r:r&r&+su#J1* B B@@ ?? ++ //r<r&ceZdZdZdZdZdZedZedZ edZ edZ ed Z ed Z ed Zd Zy )DynamicFactorMQStatesaF Helper class for describing and indexing the state vector. Parameters ---------- k_endog_M : int Number of monthly (or non-time-specific, if k_endog_Q=0) variables. k_endog_Q : int Number of quarterly variables. endog_names : list Names of the endogenous variables. factors : int, list, or dict Integer giving the number of (global) factors, a list with the names of (global) factors, or a dictionary with: - keys : names of endogenous variables - values : lists of factor names. If this is an integer, then the factor names will be 0, 1, .... factor_orders : int or dict Integer describing the order of the vector autoregression (VAR) governing all factor block dynamics or dictionary with: - keys : factor name or tuples of factor names in a block - values : integer describing the VAR order for that factor block If a dictionary, this defines the order of the factor blocks in the state vector. Otherwise, factors are ordered so that factors that load on more variables come first (and then alphabetically, to break ties). factor_multiplicities : int or dict This argument provides a convenient way to specify multiple factors that load identically on variables. For example, one may want two "global" factors (factors that load on all variables) that evolve jointly according to a VAR. One could specify two global factors in the `factors` argument and specify that they are in the same block in the `factor_orders` argument, but it is easier to specify a single global factor in the `factors` argument, and set the order in the `factor_orders` argument, and then set the factor multiplicity to 2. This argument must be an integer describing the factor multiplicity for all factors or dictionary with: - keys : factor name - values : integer describing the factor multiplicity for the factors in the given block idiosyncratic_ar1 : bool Whether or not to model the idiosyncratic component for each series as an AR(1) process. If False, the idiosyncratic component is instead modeled as white noise. Attributes ---------- k_endog : int Total number of endogenous variables. k_states : int Total number of state variables (those associated with the factors and those associated with the idiosyncratic disturbances). k_posdef : int Total number of state disturbance terms (those associated with the factors and those associated with the idiosyncratic disturbances). k_endog_M : int Number of monthly (or non-time-specific, if k_endog_Q=0) variables. k_endog_Q : int Number of quarterly variables. k_factors : int Total number of factors. Note that factor multiplicities will have already been expanded. k_states_factors : int The number of state variables associated with factors (includes both factors and lags of factors included in the state vector). k_posdef_factors : int The number of state disturbance terms associated with factors. k_states_idio : int Total number of state variables associated with idiosyncratic disturbances. k_posdef_idio : int Total number of state disturbance terms associated with idiosyncratic disturbances. k_states_idio_M : int The number of state variables associated with idiosyncratic disturbances for monthly (or non-time-specific if there are no quarterly variables) variables. If the disturbances are AR(1), then this will be equal to `k_endog_M`, otherwise it will be equal to zero. k_states_idio_Q : int The number of state variables associated with idiosyncratic disturbances for quarterly variables. This will always be equal to `k_endog_Q * 5`, even if the disturbances are not AR(1). k_posdef_idio_M : int The number of state disturbance terms associated with idiosyncratic disturbances for monthly (or non-time-specific if there are no quarterly variables) variables. If the disturbances are AR(1), then this will be equal to `k_endog_M`, otherwise it will be equal to zero. k_posdef_idio_Q : int The number of state disturbance terms associated with idiosyncratic disturbances for quarterly variables. This will always be equal to `k_endog_Q`, even if the disturbances are not AR(1). idiosyncratic_ar1 : bool Whether or not to model the idiosyncratic component for each series as an AR(1) process. factor_blocks : list of FactorBlock List of `FactorBlock` helper instances for each factor block. factor_names : list of str List of factor names. factors : dict Dictionary with: - keys : names of endogenous variables - values : lists of factor names. Note that factor multiplicities will have already been expanded. factor_orders : dict Dictionary with: - keys : tuple of factor names - values : integer describing autoregression order Note that factor multiplicities will have already been expanded. max_factor_order : int Maximum autoregression order across all factor blocks. factor_block_orders : pd.Series Series containing lag orders, with the factor block (a tuple of factor names) as the index. factor_multiplicities : dict Dictionary with: - keys : factor name - values : integer describing the factor multiplicity for the factors in the given block endog_factor_map : dict Dictionary with: - keys : endog name - values : list of factor names loading_counts : pd.Series Series containing number of endogenous variables loading on each factor, with the factor name as the index. block_loading_counts : dict Dictionary with: - keys : tuple of factor names - values : average number of endogenous variables loading on the block (note that average is over the factors in the block) Notes ----- The goal of this class is, in particular, to make it easier to retrieve indexes of subsets of the state vector. Note that the ordering of the factor blocks in the state vector is determined by the `factor_orders` argument if a dictionary. Otherwise, factors are ordered so that factors that load on more variables come first (and then alphabetically, to break ties). - `factors_L1` is an array with the indexes of first lag of the factors from each block. Ordered first by block, and then by lag. - `factors_L1_5` is an array with the indexes contains the first - fifth lags of the factors from each block. Ordered first by block, and then by lag. - `factors_L1_5_ix` is an array shaped (5, k_factors) with the indexes of the first - fifth lags of the factors from each block. - `idio_ar_L1` is an array with the indexes of the first lag of the idiosyncratic AR states, both monthly (if appliable) and quarterly. - `idio_ar_M` is a slice with the indexes of the idiosyncratic disturbance states for the monthly (or non-time-specific if there are no quarterly variables) variables. It is an empty slice if `idiosyncratic_ar1 = False`. - `idio_ar_Q` is a slice with the indexes of the idiosyncratic disturbance states and all lags, for the quarterly variables. It is an empty slice if there are no quarterly variable. - `idio_ar_Q_ix` is an array shaped (k_endog_Q, 5) with the indexes of the first - fifth lags of the idiosyncratic disturbance states for the quarterly variables. - `endog_factor_iloc` is a list of lists, with entries for each endogenous variable. The entry for variable `i`, `endog_factor_iloc[i]` is a list of indexes of the factors that variable `i` loads on. This does not include any lags, but it can be used with e.g. `factors_L1_5_ix` to get lags. c 0||_||_|j|jz|_||_t j t |tj}t|ttf} t j t |tj} |d}t j t |tj} |s| st|ts td| st|ts td| st|ts td|s| r\|r|dk(s| rt|dk(r td| r t|} nt|D cgc]} | } } |Dcic]}|| dd }}g}|jD]}|j!|t#|} | r| Dcic]}||}}| r| Dcic]}||}}|j%|||\}}||_||_||_|j-|||_|j.j0d|_|j2|jkDr&td|j2d|jd |j.j5d j7d j9j;d d gd dgj=d |_|jADcic]J}|t jB|j>jDt|d fjGd L}}tIjJ|jAd d}tIjLt|j|d jOj;d dgd dgd |_(tIjJ|jAd d}tIjLt|j|d|_)| rS|jPjA}|jRjD||_)d|jRjT_+tIjLt jXt|jRjT} |j.jZDcgc]}|| j]vr|}}t|rtIjJ|Dcgc]}|fc}d d}tIjLt j^t|t`|d}|jRjc||_)tIjLt jXt|jRjT} | je}|jgrt#| |}td|| j]|_4t jj|jR|_6|j.jD|| f|_d|_7d|_8d}g|_9|jRjuD]\} }tw| ||j.||j}|xjn|jxz c_7|xjp|j2z c_8||jxz }|jrjc||r |jnd|_=|jdz|_>|jz|j|z|_?|jr |jnd|_@|j|_A|j|jz|_B|jn|j~z|_<|jp|jz|_Cd|_Dycc} wcc}wcc}wcc}wcc}wcc}wcc}w)Nz`factors` argument must an integer number of factors, a list of global factor names, or a dictionary, mapping observed variables to factors.zC`factor_orders` argument must either be an integer or a dictionary.zK`factor_multiplicities` argument must either be an integer or a dictionary.rz+The model must contain at least one factor.zNumber of factors (zE) cannot be greater than the number of monthly endogenous variables (z).axiscountfactorFT) ascendingblock) tupleize_colsname)indexr[orderdtypezfEach factor can be assigned to at most one block of factors in `factor_orders`. Duplicate entries for r()E k_endog_Mr5k_endogidiosyncratic_ar1r> issubdtypetypeinteger isinstancelisttupledict ValueErrorr/rangevaluesextendset_apply_factor_multiplicitiesr) factor_ordersfactor_multiplicities_construct_endog_factor_mapr3shaper0sumrename reset_index sort_values set_indexloading_countskeys atleast_1dr2meanpdIndexSeriesto_frameblock_loading_countsfactor_block_ordersr\r[ concatenatecolumnstolistonesintappend duplicatedanyr.r6max_factor_orderk_states_factorsk_posdef_factors factor_blocksitemsr&r8k_states_idio_Mk_states_idio_Q k_states_idiok_posdef_idio_Mk_posdef_idio_Q k_posdef_idiok_posdef_endog_factor_iloc)r9r`r5 endog_namesr)rprqrbfactors_is_intfactors_is_list orders_is_int mult_is_intr.ir[ _factor_namesval factor_namerYrixrzmissingdefault_block_orders duplicatesduplicate_namesr4r1s r:r;zDynamicFactorMQStates.__init__@ss#"~~6 !2tG}bjjA$WtUm< d=&92::F ($% !mmD)>$?L /7D))* *M4!@9: :z*?F<= = _GqL$W): !NOO#G} 05g?11#? ?9DEt\!_,EGE >># &C   % &=) 0<>!,)-7>M> 8D%F)4&12G%G%F !%F"&!B!B ]$9"; *%:"!% @ @ [!"..44Q7 >>DNN *24>>2BC,,0NN+;2?@ @  ! ! % %1 % - 4 4W ={{GX+>6;T] + D8$ '++- / 2==##''U W(<=??Ctt| L / /XX*//1"$$&II %,,. /7%$$,HJ{{'"udm0;0==D%F!XXm((*%g N#%99 %%' ($B  ,,113D'+'?'?'C'CD'ID $29D $ $ * * / yy NN4 8 8 > >? @B $($9$9$A$A9D,"5"5"7799 w<WEkK>E(-G=B#%99RWWSWC-H35G$E ((//0DE  $99tD$<$<$B$BCDFL!,,. >> !,z":;O77F6GIJ J)//1 "t'?'? @  ! ! % %k<&? @  !" ! *.*B*B*H*H*J - &L, l $ 5 5| $0E  ! !U^^ 3 !  ! !U__ 4 ! ENN *L    % %e , -2Ct~~#~~1!11D4H4HH261G1Gt~~Q#~~!11D4H4HH--0B0BB --0B0BB #'G @E >%F@ /B9Fs+/ _5? _: _? `8A` ` `c i}|jD]`\}}g}|D]O}|j|d} | dkDr%|t| D cgc] } |d| dzc} z }?|j|Q|||<bi} |jD]q\} } t | t s| f} g}| D]D}|j|d} | dkDr%|t| D cgc] } |d| dzc} z }?||gz }F| | t |<s|| fScc} wcc} w)a Expand `factors` and `factor_orders` to account for factor multiplity. For example, if there is a `global` factor with multiplicity 2, then this method expands that into `global.1` and `global.2` in both the `factors` and `factor_orders` dictionaries. Parameters ---------- factors : dict Dictionary of {endog_name: list of factor names} factor_orders : dict Dictionary of {tuple of factor names: factor order} factor_multiplicities : dict Dictionary of {factor name: factor multiplicity} Returns ------- new_factors : dict Dictionary of {endog_name: list of factor names}, with factor names expanded to incorporate multiplicities. new_factors : dict Dictionary of {tuple of factor names: factor order}, with factor names in each tuple expanded to incorporate multiplicities. rS.)rgetrkrrfrh)r9r)rprq new_factors endog_name factors_listnew_factor_listrnrnew_factor_ordersrYr1 new_blocks r:roz2DynamicFactorMQStates._apply_factor_multiplicitiessm8 (/  6 $J  O+ 8 )--k1=q5#16q(;,-,7-qQ(@(;;O$**;7  8'6K # 6#0#6#6#8 ? E<eU+I$ / )--k1=q5+08"5&'&1M1q5'":"55I+.I  /3? eI. / ?---+(;"5s D D cg}|jD];\}}t|ttfrt |dk(s+|j |=t |rt d|dt|jt|j}t |rt d|di}|jD];\}}t|trd||<|j|Dcic]}|dc}=t|j}t |}tjtj|j |ft"tj$|dtj$|d } |jD]\}}d | j&||f<| Scc}w) as Construct mapping of observed variables to factors. Parameters ---------- factors : dict Dictionary of {endog_name: list of factor names} endog_names : list of str List of the names of the observed variables. Returns ------- endog_factor_map : pd.DataFrame Boolean dataframe with `endog_names` as the index and the factor names (computed from the `factors` input) as the columns. Each cell is True if the associated factor is allowed to load on the associated observed variable. rzyEach observed variable must be mapped to at least one factor in the `factors` dictionary. Variables missing factors are: rzwIf a `factors` dictionary is provided, then it must include entries for each observed variable. Missing variables are: r^endog)r[rWr\rT)rrfrgrhr/rrjrn differencerzstrupdater} DataFramer>zerosraboolr~r2) r9r)rrkeyvaluer.vr0r3s r:rrz1DynamicFactorMQStates._construct_endog_factor_map's,!--/ $JCedE]3s5zQs# $ w<@@GyKL Lk"--c',,..AB w<BBI!MN N !--/ ;JC%%&' U###5$9aQT$9:  ; L--/0  % << HHdllI.d ;((;W5HH\9;"--/ 4JC/3  e , 4 %:s3 F< ctj|jtfd|jD}tj |S)zFactors.c3<K|]}|jywN)r,).0rYrs r: z3DynamicFactorMQStates.factors_L1..hsJeR(()Js)r>r@rrhrr)r9ilocrs @r:r,z DynamicFactorMQStates.factors_L1ds>YYt,, -Jt7I7IJJ~~d##r<ctj|j}g}|jD]:}|j ||j j d|j<tj|dS)z3Factors plus any lags, index shaped (5, k_factors).r(rSrT) r>r@rrrr-r?r0r)r9rrrYs r:factors_L1_5_ixz%DynamicFactorMQStates.factors_L1_5_ixksmYYt,, -'' LE KK5--.66q%//J K L~~d++r<c|j}|jr||jz}n||jz}tj ||S)z2Idiosyncratic AR states, (first block / lag only).)rrbrar5r>rEr9ix1ix2s r: idio_ar_L1z DynamicFactorMQStates.idio_ar_L1tsD##  ! ! $C&CuuS~r<cz|j}|}|jr||jz }tj||S)z.Idiosyncratic AR states for monthly variables.)rrbr`r>rErs r: idio_ar_MzDynamicFactorMQStates.idio_ar_M~s;##  ! ! 4>> !CuuS~r<c|j}|jr||jz }||jdzz}tj ||S)z=Idiosyncratic AR states and all lags for quarterly variables.r()rrbr`r5r>rErs r: idio_ar_QzDynamicFactorMQStates.idio_ar_QsI##  ! ! 4>> !CDNNQ&&uuS~r<c|j}|jr||jz }|tjtj d|j zd|j fjzS)zr?r@r5rA)r9starts r: idio_ar_Q_ixz"DynamicFactorMQStates.idio_ar_Q_ixsf%%  ! ! T^^ #E  !dnn,-4>>/BDDEAF Gr<c|jag}t|jD]@}|jt j |j j|dB||_|jS)z?List of list of int, factor indexes for each observed variable.r)rrkrarr>wherer3r)r9ilocsrs r:endog_factor_ilocz'DynamicFactorMQStates.endog_factor_ilocsn  " " *E4<<( I RXXd&;&;&@&@&CDQGH I&+D #&&&r<c8|dvr t||St|)z Use square brackets to access index / slice elements. This is convenient in highlighting the indexing / slice quality of these attributes in the code below. )r,rrrrr)getattrKeyError)r9rs r: __getitem__z!DynamicFactorMQStates.__getitem__s( 0 04% %3- r<N)rJrKrLrMr;rorrrNr,rrrrrrrrOr<r:rQrQsqfl'\7.r; z$$ ,, G G ' '  r<rQceZdZdZ dfd ZedZ ddZedZ dZ d dZ edZ d!d Z d Zed Zed Zed ZdZdZfdZedZdZ d"fd Z d#dZd dZd!dZd!dZd$dZd$dZ d%fd Z d&fd Z d'fd Z d(fd Z!xZ"S))DynamicFactorMQuej Dynamic factor model with EM algorithm; option for monthly/quarterly data. Implementation of the dynamic factor model of Bańbura and Modugno (2014) ([1]_) and Bańbura, Giannone, and Reichlin (2011) ([2]_). Uses the EM algorithm for parameter fitting, and so can accommodate a large number of left-hand-side variables. Specifications can include any collection of blocks of factors, including different factor autoregression orders, and can include AR(1) processes for idiosyncratic disturbances. Can incorporate monthly/quarterly mixed frequency data along the lines of Mariano and Murasawa (2011) ([4]_). A special case of this model is the Nowcasting model of Bok et al. (2017) ([3]_). Moreover, this model can be used to compute the news associated with updated data releases. Parameters ---------- endog : array_like Observed time-series process :math:`y`. See the "Notes" section for details on how to set up a model with monthly/quarterly mixed frequency data. k_endog_monthly : int, optional If specifying a monthly/quarterly mixed frequency model in which the provided `endog` dataset contains both the monthly and quarterly data, this variable should be used to indicate how many of the variables are monthly. Note that when using the `k_endog_monthly` argument, the columns with monthly variables in `endog` should be ordered first, and the columns with quarterly variables should come afterwards. See the "Notes" section for details on how to set up a model with monthly/quarterly mixed frequency data. factors : int, list, or dict, optional Integer giving the number of (global) factors, a list with the names of (global) factors, or a dictionary with: - keys : names of endogenous variables - values : lists of factor names. If this is an integer, then the factor names will be 0, 1, .... The default is a single factor that loads on all variables. Note that there cannot be more factors specified than there are monthly variables. factor_orders : int or dict, optional Integer describing the order of the vector autoregression (VAR) governing all factor block dynamics or dictionary with: - keys : factor name or tuples of factor names in a block - values : integer describing the VAR order for that factor block If a dictionary, this defines the order of the factor blocks in the state vector. Otherwise, factors are ordered so that factors that load on more variables come first (and then alphabetically, to break ties). factor_multiplicities : int or dict, optional This argument provides a convenient way to specify multiple factors that load identically on variables. For example, one may want two "global" factors (factors that load on all variables) that evolve jointly according to a VAR. One could specify two global factors in the `factors` argument and specify that they are in the same block in the `factor_orders` argument, but it is easier to specify a single global factor in the `factors` argument, and set the order in the `factor_orders` argument, and then set the factor multiplicity to 2. This argument must be an integer describing the factor multiplicity for all factors or dictionary with: - keys : factor name - values : integer describing the factor multiplicity for the factors in the given block idiosyncratic_ar1 : bool Whether or not to model the idiosyncratic component for each series as an AR(1) process. If False, the idiosyncratic component is instead modeled as white noise. standardize : bool or tuple, optional If a boolean, whether or not to standardize each endogenous variable to have mean zero and standard deviation 1 before fitting the model. See "Notes" for details about how this option works with postestimation output. If a tuple (usually only used internally), then the tuple must have length 2, with each element containing a Pandas series with index equal to the names of the endogenous variables. The first element should contain the mean values and the second element should contain the standard deviations. Default is True. endog_quarterly : pandas.Series or pandas.DataFrame Observed quarterly variables. If provided, must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. See the "Notes" section for details on how to set up a model with monthly/quarterly mixed frequency data. init_t0 : bool, optional If True, this option initializes the Kalman filter with the distribution for :math:`\alpha_0` rather than :math:`\alpha_1`. See the "Notes" section for more details. This option is rarely used except for testing. Default is False. obs_cov_diag : bool, optional If True and if `idiosyncratic_ar1 is True`, then this option puts small positive values in the observation disturbance covariance matrix. This is not required for estimation and is rarely used except for testing. (It is sometimes used to prevent numerical errors, for example those associated with a positive semi-definite forecast error covariance matrix at the first time step when using EM initialization, but state space models in Statsmodels switch to the univariate approach in those cases, and so do not need to use this trick). Default is False. Notes ----- The basic model is: .. math:: y_t & = \Lambda f_t + \epsilon_t \\ f_t & = A_1 f_{t-1} + \dots + A_p f_{t-p} + u_t where: - :math:`y_t` is observed data at time t - :math:`\epsilon_t` is idiosyncratic disturbance at time t (see below for details, including modeling serial correlation in this term) - :math:`f_t` is the unobserved factor at time t - :math:`u_t \sim N(0, Q)` is the factor disturbance at time t and: - :math:`\Lambda` is referred to as the matrix of factor loadings - :math:`A_i` are matrices of autoregression coefficients Furthermore, we allow the idiosyncratic disturbances to be serially correlated, so that, if `idiosyncratic_ar1=True`, :math:`\epsilon_{i,t} = \rho_i \epsilon_{i,t-1} + e_{i,t}`, where :math:`e_{i,t} \sim N(0, \sigma_i^2)`. If `idiosyncratic_ar1=False`, then we instead have :math:`\epsilon_{i,t} = e_{i,t}`. This basic setup can be found in [1]_, [2]_, [3]_, and [4]_. We allow for two generalizations of this model: 1. Following [2]_, we allow multiple "blocks" of factors, which are independent from the other blocks of factors. Different blocks can be set to load on different subsets of the observed variables, and can be specified with different lag orders. 2. Following [4]_ and [2]_, we allow mixed frequency models in which both monthly and quarterly data are used. See the section on "Mixed frequency models", below, for more details. Additional notes: - The observed data may contain arbitrary patterns of missing entries. **EM algorithm** This model contains a potentially very large number of parameters, and it can be difficult and take a prohibitively long time to numerically optimize the likelihood function using quasi-Newton methods. Instead, the default fitting method in this model uses the EM algorithm, as detailed in [1]_. As a result, the model can accommodate datasets with hundreds of observed variables. **Mixed frequency data** This model can handle mixed frequency data in two ways. In this section, we only briefly describe this, and refer readers to [2]_ and [4]_ for all details. First, because there can be arbitrary patterns of missing data in the observed vector, one can simply include lower frequency variables as observed in a particular higher frequency period, and missing otherwise. For example, in a monthly model, one could include quarterly data as occurring on the third month of each quarter. To use this method, one simply needs to combine the data into a single dataset at the higher frequency that can be passed to this model as the `endog` argument. However, depending on the type of variables used in the analysis and the assumptions about the data generating process, this approach may not be valid. For example, suppose that we are interested in the growth rate of real GDP, which is measured at a quarterly frequency. If the basic factor model is specified at a monthly frequency, then the quarterly growth rate in the third month of each quarter -- which is what we actually observe -- is approximated by a particular weighted average of unobserved monthly growth rates. We need to take this particular weight moving average into account in constructing our model, and this is what the second approach does. The second approach follows [2]_ and [4]_ in constructing a state space form to explicitly model the quarterly growth rates in terms of the unobserved monthly growth rates. To use this approach, there are two methods: 1. Combine the monthly and quarterly data into a single dataset at the monthly frequency, with the monthly data in the first columns and the quarterly data in the last columns. Pass this dataset to the model as the `endog` argument and give the number of the variables that are monthly as the `k_endog_monthly` argument. 2. Construct a monthly dataset as a Pandas DataFrame with a DatetimeIndex or PeriodIndex at the monthly frequency and separately construct a quarterly dataset as a Pandas DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. Pass the monthly DataFrame to the model as the `endog` argument and pass the quarterly DataFrame to the model as the `endog_quarterly` argument. Note that this only incorporates one particular type of mixed frequency data. See also Banbura et al. (2013). "Now-Casting and the Real-Time Data Flow." for discussion about other types of mixed frequency data that are not supported by this framework. **Nowcasting and the news** Through its support for monthly/quarterly mixed frequency data, this model can allow for the nowcasting of quarterly variables based on monthly observations. In particular, [2]_ and [3]_ use this model to construct nowcasts of real GDP and analyze the impacts of "the news", derived from incoming data on a real-time basis. This latter functionality can be accessed through the `news` method of the results object. **Standardizing data** As is often the case in formulating a dynamic factor model, we do not explicitly account for the mean of each observed variable. Instead, the default behavior is to standardize each variable prior to estimation. Thus if :math:`y_t` are the given observed data, the dynamic factor model is actually estimated on the standardized data defined by: .. math:: x_{i, t} = (y_{i, t} - \bar y_i) / s_i where :math:`\bar y_i` is the sample mean and :math:`s_i` is the sample standard deviation. By default, if standardization is applied prior to estimation, results such as in-sample predictions, out-of-sample forecasts, and the computation of the "news" are reported in the scale of the original data (i.e. the model output has the reverse transformation applied before it is returned to the user). Standardization can be disabled by passing `standardization=False` to the model constructor. **Identification of factors and loadings** The estimated factors and the factor loadings in this model are only identified up to an invertible transformation. As described in (the working paper version of) [2]_, while it is possible to impose normalizations to achieve identification, the EM algorithm does will converge regardless. Moreover, for nowcasting and forecasting purposes, identification is not required. This model does not impose any normalization to identify the factors and the factor loadings. **Miscellaneous** There are two arguments available in the model constructor that are rarely used but which deserve a brief mention: `init_t0` and `obs_cov_diag`. These arguments are provided to allow exactly matching the output of other packages that have slight differences in how the underlying state space model is set up / applied. - `init_t0`: state space models in Statsmodels follow Durbin and Koopman in initializing the model with :math:`\alpha_1 \sim N(a_1, P_1)`. Other implementations sometimes initialize instead with :math:`\alpha_0 \sim N(a_0, P_0)`. We can accommodate this by prepending a row of NaNs to the observed dataset. - `obs_cov_diag`: the state space form in [1]_ incorporates non-zero (but very small) diagonal elements for the observation disturbance covariance matrix. Examples -------- Constructing and fitting a `DynamicFactorMQ` model. >>> data = sm.datasets.macrodata.load_pandas().data.iloc[-100:] >>> data.index = pd.period_range(start='1984Q4', end='2009Q3', freq='Q') >>> endog = data[['infl', 'tbilrate']].resample('M').last() >>> endog_Q = np.log(data[['realgdp', 'realcons']]).diff().iloc[1:] * 400 **Basic usage** In the simplest case, passing only the `endog` argument results in a model with a single factor that follows an AR(1) process. Note that because we are not also providing an `endog_quarterly` dataset, `endog` can be a numpy array or Pandas DataFrame with any index (it does not have to be monthly). The `summary` method can be useful in checking the model specification. >>> mod = sm.tsa.DynamicFactorMQ(endog) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 1 factors in 1 blocks # of factors: 1 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings ======================== Dep. variable 0 ------------------------ infl X tbilrate X Factor blocks: ===================== block order --------------------- 0 1 ===================== **Factors** With `factors=2`, there will be two independent factors that will each evolve according to separate AR(1) processes. >>> mod = sm.tsa.DynamicFactorMQ(endog, factors=2) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 2 factors in 2 blocks # of factors: 2 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings =================================== Dep. variable 0 1 ----------------------------------- infl X X tbilrate X X Factor blocks: ===================== block order --------------------- 0 1 1 1 ===================== **Factor multiplicities** By instead specifying `factor_multiplicities=2`, we would still have two factors, but they would be dependent and would evolve jointly according to a VAR(1) process. >>> mod = sm.tsa.DynamicFactorMQ(endog, factor_multiplicities=2) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 2 factors in 1 blocks # of factors: 2 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings =================================== Dep. variable 0.1 0.2 ----------------------------------- infl X X tbilrate X X Factor blocks: ===================== block order --------------------- 0.1, 0.2 1 ===================== **Factor orders** In either of the above cases, we could extend the order of the (vector) autoregressions by using the `factor_orders` argument. For example, the below model would contain two independent factors that each evolve according to a separate AR(2) process: >>> mod = sm.tsa.DynamicFactorMQ(endog, factors=2, factor_orders=2) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 2 factors in 2 blocks # of factors: 2 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings =================================== Dep. variable 0 1 ----------------------------------- infl X X tbilrate X X Factor blocks: ===================== block order --------------------- 0 2 1 2 ===================== **Serial correlation in the idiosyncratic disturbances** By default, the model allows each idiosyncratic disturbance terms to evolve according to an AR(1) process. If preferred, they can instead be specified to be serially independent by passing `ididosyncratic_ar1=False`. >>> mod = sm.tsa.DynamicFactorMQ(endog, idiosyncratic_ar1=False) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 1 factors in 1 blocks # of factors: 1 + iid idiosyncratic Idiosyncratic disturbances: iid Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings ======================== Dep. variable 0 ------------------------ infl X tbilrate X Factor blocks: ===================== block order --------------------- 0 1 ===================== *Monthly / Quarterly mixed frequency* To specify a monthly / quarterly mixed frequency model see the (Notes section for more details about these models): >>> mod = sm.tsa.DynamicFactorMQ(endog, endog_quarterly=endog_Q) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 1 factors in 1 blocks # of quarterly variables: 2 + Mixed frequency (M/Q) # of factors: 1 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings ======================== Dep. variable 0 ------------------------ infl X tbilrate X realgdp X realcons X Factor blocks: ===================== block order --------------------- 0 1 ===================== *Customize observed variable / factor loadings* To specify that certain that certain observed variables only load on certain factors, it is possible to pass a dictionary to the `factors` argument. >>> factors = {'infl': ['global'] ... 'tbilrate': ['global'] ... 'realgdp': ['global', 'real'] ... 'realcons': ['global', 'real']} >>> mod = sm.tsa.DynamicFactorMQ(endog, endog_quarterly=endog_Q) >>> print(mod.summary()) Model Specification: Dynamic Factor Model ========================================================================== Model: Dynamic Factor Model # of monthly variables: 2 + 2 factors in 2 blocks # of quarterly variables: 2 + Mixed frequency (M/Q) # of factor blocks: 2 + AR(1) idiosyncratic Idiosyncratic disturbances: AR(1) Sample: 1984-10 Standardize variables: True - 2009-09 Observed variables / factor loadings =================================== Dep. variable global real ----------------------------------- infl X tbilrate X realgdp X X realcons X X Factor blocks: ===================== block order --------------------- global 1 real 1 ===================== **Fitting parameters** To fit the model, use the `fit` method. This method uses the EM algorithm by default. >>> mod = sm.tsa.DynamicFactorMQ(endog) >>> res = mod.fit() >>> print(res.summary()) Dynamic Factor Results ========================================================================== Dep. Variable: ['infl', 'tbilrate'] No. Observations: 300 Model: Dynamic Factor Model Log Likelihood -127.909 + 1 factors in 1 blocks AIC 271.817 + AR(1) idiosyncratic BIC 301.447 Date: Tue, 04 Aug 2020 HQIC 283.675 Time: 15:59:11 EM Iterations 83 Sample: 10-31-1984 - 09-30-2009 Covariance Type: Not computed Observation equation: ============================================================== Factor loadings: 0 idiosyncratic: AR(1) var. -------------------------------------------------------------- infl -0.67 0.39 0.73 tbilrate -0.63 0.99 0.01 Transition: Factor block 0 ======================================= L1.0 error variance --------------------------------------- 0 0.98 0.01 ======================================= Warnings: [1] Covariance matrix not calculated. *Displaying iteration progress* To display information about the EM iterations, use the `disp` argument. >>> mod = sm.tsa.DynamicFactorMQ(endog) >>> res = mod.fit(disp=10) EM start iterations, llf=-291.21 EM iteration 10, llf=-157.17, convergence criterion=0.053801 EM iteration 20, llf=-128.99, convergence criterion=0.0035545 EM iteration 30, llf=-127.97, convergence criterion=0.00010224 EM iteration 40, llf=-127.93, convergence criterion=1.3281e-05 EM iteration 50, llf=-127.92, convergence criterion=5.4725e-06 EM iteration 60, llf=-127.91, convergence criterion=2.8665e-06 EM iteration 70, llf=-127.91, convergence criterion=1.6999e-06 EM iteration 80, llf=-127.91, convergence criterion=1.1085e-06 EM converged at iteration 83, llf=-127.91, convergence criterion=9.9004e-07 < tolerance=1e-06 **Results: forecasting, impulse responses, and more** One the model is fitted, there are a number of methods available from the results object. Some examples include: *Forecasting* >>> mod = sm.tsa.DynamicFactorMQ(endog) >>> res = mod.fit() >>> print(res.forecast(steps=5)) infl tbilrate 2009-10 1.784169 0.260401 2009-11 1.735848 0.305981 2009-12 1.730674 0.350968 2010-01 1.742110 0.395369 2010-02 1.759786 0.439194 *Impulse responses* >>> mod = sm.tsa.DynamicFactorMQ(endog) >>> res = mod.fit() >>> print(res.impulse_responses(steps=5)) infl tbilrate 0 -1.511956 -1.341498 1 -1.483172 -1.315960 2 -1.454937 -1.290908 3 -1.427240 -1.266333 4 -1.400069 -1.242226 5 -1.373416 -1.218578 For other available methods (including in-sample prediction, simulation of time series, extending the results to incorporate new data, and the news), see the documentation for state space models. References ---------- .. [1] Bańbura, Marta, and Michele Modugno. "Maximum likelihood estimation of factor models on datasets with arbitrary pattern of missing data." Journal of Applied Econometrics 29, no. 1 (2014): 133-160. .. [2] Bańbura, Marta, Domenico Giannone, and Lucrezia Reichlin. "Nowcasting." The Oxford Handbook of Economic Forecasting. July 8, 2011. .. [3] Bok, Brandyn, Daniele Caratelli, Domenico Giannone, Argia M. Sbordone, and Andrea Tambalotti. 2018. "Macroeconomic Nowcasting and Forecasting with Big Data." Annual Review of Economics 10 (1): 615-43. https://doi.org/10.1146/annurev-economics-080217-053214. .. [4] Mariano, Roberto S., and Yasutomo Murasawa. "A coincident index, common factors, and monthly real GDP." Oxford Bulletin of Economics and Statistics 72, no. 1 (2010): 27-46. c P|"| td|j||\}}t|d} | r+t|tj rH|j }n7tj|dkrtj|j}||jd}| r|jj} nA|jddk(rdg} n+t|jdDcgc] }d|dz } }t|d|_|jd|j z |_t%|j |j"| ||||x}|_||_||_||_|j&j.|_|j&j0|_|j&j2|_|j&j4|_t7|j0|_|j&j:|_||_| |_| |_ |j>r| r[t jB|jDddz |jDd|jDjF}|jI|}nHtjJtjLg|jdz|jfj}t|tNrt7|dk(rz|\}}|jd}t|tj rJ|jDjQt jR| std | d |jDd tjT|}t|tj rJ|jDjQt jR| std | d |jDd tjT|}tj||fk7stj||fk7rtd |d d}| rt j || }t j || }n}|dvr%|jWd}|jYd}nT|dvrEtjZ|jd}tj\|jd}n td||_/||_0||_1tjd|j`dkrZtjf|j`dk}tjh| |dj}td|d |jbr||j^z |j`z }tjjd|j tjj|j ddx}|_6to||f|jr|jtd| |jbr?|jvjx|j`z|j^z|jv_<d| vr)|jzj}|j|j<r+tj|j |d|d|df<gd}tt7|D]<}||}|tj|j"z|d|d|ddd|ff<>|j@r%tj|jdz|d <|jD]r}|j4dk(rd}n+tjZ|j4|j4f}tdg|g|jzzj|d!|d"|d"f<t|j"dk(rd}n+tjZ|j"|j"f}tdg|gd#zzj|d!|d$|d$f<dx}}|jD]D}||j4z }tj|j4|d%|d&dddf||f<|}F|j<r;||j z}tj|j |d%|d||f<|}||j"z}tj|j"|d%|ddddf||f<td'tj|j.jfd(tj|jDcgc]}|j4dz|jz c}fd)tj|jDcgc]!}|j4|j4dzzdz#c}fd*|j<r |jndfd+|jfg|_Itjt|jj|_Ktjtj|jtjt|jjdd}tt|jj||_Ri|_S|xjgd,t| jzz c_Tycc}wcc}wcc}w)-NzIf `endog_quarterly` is specified, then `endog` must contain only monthly variables, and so `k_endog_monthly` cannot be specified since it will be inferred from the shape of `endog`.rSyk_endog_monthlyr)freqzLInvalid value passed for `standardize`: if a Pandas Series, must have index z. Got rzEInvalid value passed for `standardize`: each element must be shaped (z,).Tr\)rSTrT)rFz'Invalid value passed for `standardize`.绽|=zwConstant variable(s) found in observed variables, but constants cannot be included in this model. These variables are: )MQ)r8rrdesignrrrSrrrSrr-C6?obs_cov transitionr)r(r selectionr+loadings factor_ar factor_covrbidiosyncratic_var)r)rprqrb standardizeinit_t0 obs_cov_diag)Urjconstruct_endogr rfr}rrr>ndim atleast_2drArsrrrkr r`r5rQ_sr)rprqr3rr.r0r/k_factor_blocksrrbrr period_ranger\rreindexc_nanrhequalsr~r{r|stdrr _endog_mean _endog_stdrrrarrayrE_osuperr;r8rdata orig_endogssm initialize_default_initializationeyerarrr7rrtrlr1paramsrgk_paramssplitr@cumsumriziprz_p_loading_constraints _init_keys)r9rrr)rprqrbrendog_quarterlyrrkwargsendog_is_pandasrrsr endog_mean endog_stdrnamesrC multipliersmrYtmprr __class__s r:r;zDynamicFactorMQ.__init__s  &* ":;; &*%9%9&( "E?*5$7 %+(wwu~! e,..  "#kk!nO --..0K{{1~""e 49%++a.4IJq1q5'{J J!/3DEQ$..8, NNDNNK 02CE EDG  *%:" $ 8 8#'77#>#>  GG00**"4#;#;< $ 8 8!2 ( <<__U[[^a%7R*/++*:*:< b)rvvhQ7@ACC k5 )c+.>!.C$/ !J  AA:ryy1$$++BHH[,AB "%%0M 8H8H7I"LMM ]]:6 9bii0OO**288K+@A "%%0M 7H"KLLMM)4 $,0Ct0K "==>Cs"DEEKYYzE IIi{C I %+J q )I J &%++a.1J A/IFG G%#& 66$//E) *$//E12BHH[)"Q%0779EEEJG1NO O   T---@E't~~') )DG  #ajj #! #    $$t69I9II II  6 ) HH   < < > ?  ! !57VVDNN5KD1S61[>1 2% s;'( ,AAABFF4>>** 1S61^#4QT#:: ; ,    ffT\\2T9DO__ GE!#hhAB !uu/B/B'B!BCEE uY/y1AA B  G >>Q C((DNNDNN;1[> 9: c__ E 5?? "Cu' eL1!Q$7S@ AC    ! !&C9;9ODa nc#g5 6CDNN" FF4>> " [!N+AqD13s7 :;"  5 5 < <= > "&&/0"@&+#(//1"4u7I7I"I"@A B 26601#A',$)??eoo6I#Ja#O#AB C $ 6 6 A ? $,, /#12 tDKK$6$6$89: XXbii .iiT[[%7%7%9 :;CR@Bs4;;++-r23%'! #6;;=12 2aK|"@#As4l!#l)&l#cb|d}t|tjtjfst d|z|j j dvrt d|zt|j ddddk(s)t|j dd }t d |d |zd }t|tjtjfst d |z|j j dvrt d|zt|j ddddk(s)t|j dd }t d|d |zt|j dr|jd}t|j dr|jd}|j}|j j|_|jtj}|jtj}|j j|_tj ||gd}|j"j%}|j'dkDr|j"j(j+t,}|j D]?} |j.| } | dk(r|| k(} t1| D cgc] } | | dz c} || <A||_n|j}|j2} t5| dk(r| dnd}||fScc} w)aR Construct a combined dataset from separate monthly and quarterly data. Parameters ---------- endog_monthly : array_like Monthly dataset. If a quarterly dataset is given, then this must be a Pandas object with a PeriodIndex or DatetimeIndex at a monthly frequency. endog_quarterly : array_like or None Quarterly dataset. If not None, then this must be a Pandas object with a PeriodIndex or DatetimeIndex at a quarterly frequency. Returns ------- endog : array_like If both endog_monthly and endog_quarterly were given, this is a Pandas DataFrame with a PeriodIndex at the monthly frequency, with all of the columns from `endog_monthly` ordered first and the columns from `endog_quarterly` ordered afterwards. Otherwise it is simply the input `endog_monthly` dataset. k_endog_monthly : int The number of monthly variables (which are ordered first) in the returned `endog` dataset. zIf given both monthly and quarterly data then the monthly dataset must be a Pandas object with a date index at a monthly frequency.z.Given monthly dataset is not a Pandas object. ) datetime64periodz9Given monthly dataset has an index with non-date values. freqstrNrrNonezIndex of given monthly dataset has a non-monthly frequency (to check this, examine the `freqstr` attribute of the index of the dataset - it should start with M if it is monthly). Got z. zlIf a quarterly dataset is given, then it must be a Pandas object with a date index at a quarterly frequency.z0Given quarterly dataset is not a Pandas object. z;Given quarterly dataset has an index with non-date values. rzIndex of given quarterly dataset has a non-quarterly frequency (to check this, examine the `freqstr` attribute of the index of the dataset - it should start with Q if it is quarterly). Got to_periodrSrTr)rfr}rrrjr\ inferred_typerhasattrr!copy to_timestampresamplerfirstrconcatr value_countsr6rlastypeobjectr2rkrsr/)cls endog_monthlyrbase_msgrquarterly_resampr column_countsrr[rVmaskrrsrs r:rzDynamicFactorMQ.construct_endogsd8  &LHmbii-FG "46>"?@@$$22;EE "ACK"LMM]00)SA!DK!-"5"5y&I ") *1  "5 8@ "@AA&Ho 2<"?@@ &&44=GG "ACK"LMM_22IsCAF#M!/"7"7FK ") *1  "5 8@ "@AA}**K8 - 7 7 < ,,k:"1";";C"@ /335 %5%;%;%H%H%J  "/88EKKM /88CIIK %5%;%;%E%E%G  "II}.>?aHE"MM668M  "Q&--..55f=)//MD)--d3Ez "d?D?DU|$L!va!eW%5$LGDM M !( !&&(E##&)%jAo%(1o%%%MsL,c |r'|jr|j|jf|d<|j|f||d|}|S)a Clone state space model with new data and optionally new specification. Parameters ---------- endog : array_like The observed time-series process :math:`y` k_endog_monthly : int, optional If specifying a monthly/quarterly mixed frequency model in which the provided `endog` dataset contains both the monthly and quarterly data, this variable should be used to indicate how many of the variables are monthly. endog_quarterly : array_like, optional Observations of quarterly variables. If provided, must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. kwargs Keyword arguments to pass to the new model class to change the model specification. Returns ------- model : DynamicFactorMQ instance r)rr)rrr_clone_from_init_kwds)r9rrrretain_standardizationrmods r:clonezDynamicFactorMQ.clone>sW4 "d&6&6%)%5%5t$GF= !(d(( 7#2+7/57 r<c2dttjfiS)Nfit)DynamicFactorMQResultsrMLEResultsWrapperr9s r: _res_classeszDynamicFactorMQ._res_classes_s.0J0JKLLr<cd|j}tj|j}|jD]}|j |dd|j r=t|dj|djD]}|j |d|j |dd|S)Nr) stationaryrr) rrInitializationr8rrnrbrkrstop)r9rinitrYrs r:rz'DynamicFactorMQ._default_initializationcs GG,,T]];__ 5E HHU9%| 4 5  ! !1[>//;1D1DE *L) * ;. r<cF||dus|jdk(rdnd}|dur|dur td||du}|j}t|ts|g}|r|Dcgc] }t |}}|dur'|}|Dcgc]}t ||kr|n|d|dz}}|Scc}wcc}w)NFrSz?Can only truncate endog names if they are returned as a string.z...)rarjrrfrgrr/)r9truncate as_stringrr[rs r:_get_endog_namesz DynamicFactorMQ._get_endog_namesns   )U 2dlla6GuRH  (%"7:; ;   -I&& +t,&-K 1<=3t9=K= 5 A'24#$'t9>4tBQx%7GG4K4>4s B6Bcd|jd|jdg}|jdkDr|jd|jrdnd}|j|d|S) NzDynamic Factor Modelz factors in z blocksrzMixed frequency (M/Q)AR(1)iidz idiosyncratic)r0rr5rrb)r9 model_name error_types r: _model_namezDynamicFactorMQ._model_namesr #~~l4+?+?*@ HJ  >>A    5 6 $ 6 6WE ZL78r<cD|j|d}d}|jr)|j}|d}d|zg}|d}|dd|zzgz }n#tddt|jzg}|j }g}|j d |dgftd t|D]} |j d d || zgf|d |dgfd |d gfgz }g} |jdkDr"| d|jgfd|jgfgz } n| d|jgfgz } |jd k(r| d|jgfgz } n| d|jgfgz } | d|jrdndgfd|jgfgz } t!} ||_| j%||| |d } |`|j&j)dd d} || _ | j-}|D]\}}| |dt|dzzz| |<!d| j*_| j5} | j6}| j8j;tj=}d}d}t?|||t@|}| jBjE| || d z } |jFj5} | d j;d!| d < | d"gj;t| d"g<| j6}| j8j;tj=}d}d#}t?|||t@|}| jBjE| || d z } | S#t.$r| j1}YwxYw#t.$r | d"gjIt| d"g<YwxYw)$a` Create a summary table describing the model. Parameters ---------- truncate_endog_names : int, optional The number of characters to show for names of observed variables. Default is 24 if there is more than one observed variable, or an unlimited number of there is only one. T)rDrEz)Model Specification: Dynamic Factor Modelrz%srz- z - zModel:rSz+ zSample:z# of monthly variables:z# of quarterly variables:z# of observed variables:z # of factors:z# of factor blocks:zIdiosyncratic disturbances:rHrIzStandardize variables:)gleftgrighttitleX)TF rz Dep. variableNz$Observed variables / factor loadingstxt_fmtrQrYc$dj|S)Nz, )join)r.s r:z)DynamicFactorMQ.summary..s>A  *T^^,<=,t~~.>?A AI 5 ~FG GI   1 $ ?T^^,<=> >I 043G3G2HIJ JI4#'#9#9uEG/$2B2B1CDF F ) HY&+  - J$$,,C-CD   %JJLE ?ID#dsc$i1n'=>DJ ?) !kk  ((-446  6  e- h.A ''335W )) 8:W  <"G9o11#6D'Okk  ((-446     e- h.A S %NN$E %4 <"G9o66s;D'O ""r<c \g}|jjD]a}||jddDcgc]}|c}z }td|jD]%}||jDcgc] }d|d| c}z }'c|j }|j r(||jd}||Dcgc]}d| c}z }||jd}||Dcgc]}d| c}z }tdd D]}||Dcgc] }d|d | c}z }|Scc}wcc}wcc}wcc}wcc}w) zA(list of str) List of human readable names for unobserved states.NrSLrrzeps_M.rzeps_Q.r(z.eps_Q.)rrr.rkr7rFrbr)r9 state_namesrYr[rr endog_names_M endog_names_Qs r:ruzDynamicFactorMQ.state_namessj WW** @E %2D2DQ2GH$tfIH HK1e112 @,1,>,> @$(#$A3av @@  @ @++-  ! !' 5M  FfTFOF FK#DGGCL1  MBD&BB q! LA ]KTas'$0K KK L!I @GCKs D%D 0 D D$=D)cg}|jd}|D]a}|jjD]F}|jD]5}|jj ||fs|j d|d|7Hc|jjD]Y}|jD]H}|td|jdzDcgc]}|jD] }d|d|d| c}}z }J[tt|jjD]f}|jj|}|t|jD cgc]'} t| dzD]} d|d | dzd | dzd )c} } z }h|jrP||jd } || D cgc]} d |  c} z }||jd} || D cgc]} d|  c} z }||D cgc]} d|  c} z }|Scc}}wcc} } wcc} wcc} wcc} w)z5(list of str) List of human readable parameter names.FrEzloading.z->rSrtrzfb(z ).cov.chol[,]rz L1.eps_M.rz L1.eps_Q.zsigma2.) rFrrr.r3r2rrkr1r/r0rbr)r9 param_namesrrrYr to_factorr from_factorjkrvr[rws r:r|zDynamicFactorMQ.param_namessr ++e+< % DJ.. D#(#5#5DK,,00[1HI#**&{m2j\BDD D DWW** GE"// G ).q%2D2Dq2H)I G$%383E3E G$/#$A3a }Byk B G B GG  G Gs4770012 3AGG))!,E %*5??%;3 !%*1q5\3 !"!KAwaAwa@3@3 3K 3  ! !' 5M =I4iv.I IK' 5M =I4iv.I IK [AT'$(AA - G3JJBs#G2 ,G8 ! G> H Hc xtj|jtj}|jj d|j }g}tjtj|jjjd}|jD]}tj|jdd|fd}t!|dk(r3tj|jjdd|fd}|dd|f}t#|ddd }|j%|j&|dd|fxx|j(zcc<tj*|d }g} g} t-|j D]} |j.j0| } |dd| f} t3|jdd| f| d }|j5}| |j6j9z } | j%|j:t-|j |j<D]} |j.j0| } t?|dd| fd d} tA|jdd| f| d }|jC|jE| }| |j6dt!| j9z } | j%|jF| ||jHd<d}g}g}d} |j.jJD]}|dd| | |jLzf}| |jLz } |jNdk(r;|jLdk(r@tQ||jNddf}|jR}|dd}|dd}|jRdd}n|jLdkDrtU|}|j5|jNdd}|j6jVjY}tjZj]|j^}|tj`|}tjbtjd||jL|jL|jNfd}tgdgti z}|setkd|jddddtjl|jod }|tjp|jL}|j9z }|j9z }|||jHd<|||jHd<|jrrg}g} t-|j D]} tQ| | dd}!|!jR}|j%tjt|ddd | j%tjt|dd!tjvt-|j |j<D]} |jdd| fjy}"| | |"tjz|"<t}|"}!|!j5d"dd#}#|!j|#d$d%}#|j%tjt|#ddd | j%tjt|#dd!tjv|||jHd&<| ||jHd'<|St-|j D cgc]} tj| | } } t-|j |j<D]} |jdd| fjy}"| | |"tjz|"<t}|"}!|!j5dd(}#| j%tjt|#dd!tjv| ||jHd'<|Scc} w))z>(array) Starting parameters for maximum likelihood estimation.r^NW) requirementsrrSeigF)ncompmethod normalizerTdrop)exogrin)originalrrT)r]rr)maxlagsictrend)rrrSzPNon-stationary starting factor autoregressive parameters found for factor block z%. Using zeros as starting parameters.rr)rSrrc)r]rgGzgGz?gh㈵> )maxiter return_paramsdispr()rrrbr)rr)Ar>rr float64r3rr`requirer}rr interpolatebfillr.rr2r/rrr) projectionrrkrrr r8rrresidrarrfit_constrainedloading_constraintsresid_responser rr0r1r start_paramsrrAravellinalgcholeskysigma_utril_indices_from transposer?rrgrdiagr tril_indicesrbclipinfr$isnanrfit_emvar)$r9rendog_factor_map_Mr)rr[endog_ix factor_endogres_pcarrr factor_ix factor_exogmod_olsres_olsmod_glmres_glmr>rrrY factors_endog mod_factorsspblock_factor_arblock_factor_covcoefficient_matrices res_factorsrt cov_factoridio_ar1idio_varmod_idiorres_idios$ r:rzDynamicFactorMQ.start_paramsFs[$--rzz: "2277H LL $ 0 0 2 8 8 : %% 5Dxx 2 6 6q$w ?@CH8}!88D$9$9$=$=ag$FGJ H-L,aOG NN7?? + !X+ '"4"4 4  5..q1 t~~& (A11!4I!!Y,/K$**QT*fMGkkmG --/ /H LL '  (t~~t||4 1A11!4I I!6DIK$**QT*KHG--d.F.Fq.IJG I7>>@ @H LL// 0  1'/twwz"#    WW**+ 4E#AqU__)<'<$<=M  A!!Q&!#%m-2-?-?A,FH  --"$Sb'#%bc7 '2'?'?'D$1$!-0 )oo!..4s.D #."4"4"6"6"<"<">II&&{':':;#$R%9%9!%<#= ')||JJ % % 2 2 456?(A$ 'sT3G2G-H'HIJ++,-$$%&'"WW]%6%6A%6%>? ru?@! //1 1I *113 3JW+ 4X(1tww{#$(2tww|$%  ! !HH4>>* ?"589CH**1ud ;<2bff =>  ? 4>>4<<8 DJJq!t$))+"'(288A;,'?#<<$-2(4#??8Q9=+? UD AB T266 BC D4>4<<8 DJJq!t$))+"'(288A;,'?#<r?rrrrrbrkrar) r9 unconstrained constrainedunconstrained_factor_arconstrained_factor_arrrYlength tmp_coefftmp_cov_rs r:transform_paramsz DynamicFactorMQ.transform_paramss&$((* #0 0D"E " WW** E__a'%*<*<r?rrrrrbrkrar) r9rrrrrrYrrrrrs r:untransform_paramsz"DynamicFactorMQ.untransform_paramss$$((* !,DGGK,@ A"$ WW** E__a'%*<*<y?FHLIq #y'8'?'?'A A # KA /F dggk*+  ! !"477+>#?@Ht||,;.2(1QU2CDQG;.M$''"56 7 $''"56 7 < dgg123;.rc t||fi|}|j}|j}|j}||d}d}t |j D]A}|jj|} t| } |d| } |||| z|d|| f<|| z }Ctjgddddf} t |j |jD]i}|jj|} t| } |ddd| f} tj|||| z| z|d|| jf<|| z }k||d} d}|jD]p}|jd z|jz}tj | |||z|j|j|jzf}||z }||d |d|d f<r||d }d}d}|jD]}|j|jd zzd z}tj"|j|jf|j$}||||z|tj&|<||z }||j(z}||jz}||d||||f<|}|j*r)tj,||d}||d |d|df<|j*r:tj,||d|d|jd|jdf<y||d}tj,||d|d|d|df<tj,||d|d|jd|jdf<y)aK Update the parameters of the model. Parameters ---------- params : array_like Array of new parameters. transformed : bool, optional Whether or not `params` is already transformed. If set to False, `transform_params` is called. Default is True. rrr,rrNrrrrr*rrSr^ state_covrbrrrrr)rrrrr rkr`rr/r>rrarrr0r1r?rr_rrArbr)r9rrrCrprrrrr0rrrrYr Arrrtrralpharrs r:rzDynamicFactorMQ.update5s1&1 GG GG GG!J-(t~~& A77,,Q/DD I,-I+3E%):K+LD1i' ( Y E   hh/48 t~~t||4 A77,,Q/DD I+,QW5I3588uy01K?4AD1ioo// 0 Y E  1[>* __ ME)E,>,>>H % 01%//E4F4F"FGIA X EKLDu\2E,4GG H  MAlO, __ E%//A*=>!CH%//5??;%||-A)3E%(:J)KAb""1% & X EACCA'C23Dc#gs3w. /C   ! !GGF1%8#9:;ECHDq,? @  ! !q!4567 dnnot~~> ?a 345H.0gghqv6F.GDAcFAcF* +3() dnnot~~> ?r<cddtj|jz jztjdtj zzS)z Constant term in the joint log-likelihood function. Useful in facilitating comparisons to other packages that exclude the constant from the log-likelihood computation. grSr)r>rrrtlogpir;s r:loglike_constantz DynamicFactorMQ.loglike_constants@q288DJJ//4466BEE 9JJJr<cj||jkr td||jvr|jj|j }t j|dz|dzf}t j|jd}t jgd}t j|ddt j|dzj|dz|f|ddd|f<t jd g|dzz|dd|df<||f|j|<|j|S) al Matrix formulation of quarterly variables' factor loading constraints. Parameters ---------- i : int Index of the `endog` variable to compute constraints for. Returns ------- R : array (k_constraints, k_factors * 5) q : array (k_constraints,) Notes ----- If the factors were known, then the factor loadings for the ith quarterly variable would be computed by a linear regression of the form y_i = A_i' f + B_i' L1.f + C_i' L2.f + D_i' L3.f + E_i' L4.f where: - f is (k_i x 1) and collects all of the factors that load on y_i - L{j}.f is (k_i x 1) and collects the jth lag of each factor - A_i, ..., E_i are (k_i x 1) and collect factor loadings As the observed variable is quarterly while the factors are monthly, we want to restrict the estimated regression coefficients to be: y_i = A_i f + 2 A_i L1.f + 3 A_i L2.f + 2 A_i L3.f + A_i L4.f Stack the unconstrained coefficients: \Lambda_i = [A_i' B_i' ... E_i']' Then the constraints can be written as follows, for l = 1, ..., k_i - 2 A_{i,l} - B_{i,l} = 0 - 3 A_{i,l} - C_{i,l} = 0 - 2 A_{i,l} - D_{i,l} = 0 - A_{i,l} - E_{i,l} = 0 So that k_constraints = 4 * k_i. In matrix form the constraints are: .. math:: R \Lambda_i = q where :math:`\Lambda_i` is shaped `(k_i * 5,)`, :math:`R` is shaped `(k_constraints, k_i * 5)`, and :math:`q` is shaped `(k_constraints,)`. For example, for the case that k_i = 2, we can write: | 2 0 -1 0 0 0 0 0 0 0 | | A_{i,1} | | 0 | | 0 2 0 -1 0 0 0 0 0 0 | | A_{i,2} | | 0 | | 3 0 0 0 -1 0 0 0 0 0 | | B_{i,1} | | 0 | | 0 3 0 0 0 -1 0 0 0 0 | | B_{i,2} | | 0 | | 2 0 0 0 0 0 -1 0 0 0 | | C_{i,1} | = | 0 | | 0 2 0 0 0 0 0 -1 0 0 | | C_{i,2} | | 0 | | 1 0 0 0 0 0 0 0 -1 0 | | D_{i,1} | | 0 | | 0 1 0 0 0 0 0 0 0 -1 | | D_{i,2} | | 0 | | E_{i,1} | | 0 | | E_{i,2} | | 0 | z%No constraints for monthly variables.rr(rrrSN.Nr)r`rjrr3rrtr>rrsrr?rrAr)r9rr0Rqrs r:rz#DynamicFactorMQ.loading_constraintssB t~~ DE E D-- ---221599;I)a-Q78A$A((?3K!zzQR266)#4Y#??BBQ * ,Aa)m  "wwty1}'=>Aam ,-q6D % %a (((++r<c |dk(r!|jd||||||| | | | ||||d|St|did|d|d|d|d|d|d |d | d | d | d |d|d|d|d|d||S)a Fits the model by maximum likelihood via Kalman filter. Parameters ---------- start_params : array_like, optional Initial guess of the solution for the loglikelihood maximization. If None, the default is given by Model.start_params. transformed : bool, optional Whether or not `start_params` is already transformed. Default is True. includes_fixed : bool, optional If parameters were previously fixed with the `fix_params` method, this argument describes whether or not `start_params` also includes the fixed parameters, in addition to the free parameters. Default is False. cov_type : str, optional The `cov_type` keyword governs the method for calculating the covariance matrix of parameter estimates. Can be one of: - 'opg' for the outer product of gradient estimator - 'oim' for the observed information matrix estimator, calculated using the method of Harvey (1989) - 'approx' for the observed information matrix estimator, calculated using a numerical approximation of the Hessian matrix. - 'robust' for an approximate (quasi-maximum likelihood) covariance matrix that may be valid even in the presence of some misspecifications. Intermediate calculations use the 'oim' method. - 'robust_approx' is the same as 'robust' except that the intermediate calculations use the 'approx' method. - 'none' for no covariance matrix calculation. Default is 'none', since computing this matrix can be very slow when there are a large number of parameters. cov_kwds : dict or None, optional A dictionary of arguments affecting covariance matrix computation. **opg, oim, approx, robust, robust_approx** - 'approx_complex_step' : bool, optional - If True, numerical approximations are computed using complex-step methods. If False, numerical approximations are computed using finite difference methods. Default is True. - 'approx_centered' : bool, optional - If True, numerical approximations computed using finite difference methods use a centered approximation. Default is False. method : str, optional The `method` determines which solver from `scipy.optimize` is used, and it can be chosen from among the following strings: - 'em' for the EM algorithm - 'newton' for Newton-Raphson - 'nm' for Nelder-Mead - 'bfgs' for Broyden-Fletcher-Goldfarb-Shanno (BFGS) - 'lbfgs' for limited-memory BFGS with optional box constraints - 'powell' for modified Powell's method - 'cg' for conjugate gradient - 'ncg' for Newton-conjugate gradient - 'basinhopping' for global basin-hopping solver The explicit arguments in `fit` are passed to the solver, with the exception of the basin-hopping solver. Each solver has several optional arguments that are not the same across solvers. See the notes section below (or scipy.optimize) for the available arguments and for the list of explicit arguments that the basin-hopping solver supports. maxiter : int, optional The maximum number of iterations to perform. tolerance : float, optional Tolerance to use for convergence checking when using the EM algorithm. To set the tolerance for other methods, pass the optimizer-specific keyword argument(s). full_output : bool, optional Set to True to have all available output in the Results object's mle_retvals attribute. The output is dependent on the solver. See LikelihoodModelResults notes section for more information. disp : bool, optional Set to True to print convergence messages. callback : callable callback(xk), optional Called after each iteration, as callback(xk), where xk is the current parameter vector. return_params : bool, optional Whether or not to return only the array of maximizing parameters. Default is False. optim_score : {'harvey', 'approx'} or None, optional The method by which the score vector is calculated. 'harvey' uses the method from Harvey (1989), 'approx' uses either finite difference or complex step differentiation depending upon the value of `optim_complex_step`, and None uses the built-in gradient approximation of the optimizer. Default is None. This keyword is only relevant if the optimization method uses the score. optim_complex_step : bool, optional Whether or not to use complex step differentiation when approximating the score; if False, finite difference approximation is used. Default is True. This keyword is only relevant if `optim_score` is set to 'harvey' or 'approx'. optim_hessian : {'opg','oim','approx'}, optional The method by which the Hessian is numerically approximated. 'opg' uses outer product of gradients, 'oim' uses the information matrix formula from Harvey (1989), and 'approx' uses numerical approximation. This keyword is only relevant if the optimization method uses the Hessian matrix. low_memory : bool, optional If set to True, techniques are applied to substantially reduce memory usage. If used, some features of the results object will not be available (including smoothed results and in-sample prediction), although out-of-sample forecasting is possible. Note that this option is not available when using the EM algorithm (which is the default for this model). Default is False. llf_decrease_action : {'ignore', 'warn', 'revert'}, optional Action to take if the log-likelihood decreases in an EM iteration. 'ignore' continues the iterations, 'warn' issues a warning but continues the iterations, while 'revert' ends the iterations and returns the result from the last good iteration. Default is 'warn'. llf_decrease_tolerance : float, optional Minimum size of the log-likelihood decrease required to trigger a warning or to end the EM iterations. Setting this value slightly larger than zero allows small decreases in the log-likelihood that may be caused by numerical issues. If set to zero, then any decrease will trigger the `llf_decrease_action`. Default is 1e-4. **kwargs Additional keyword arguments to pass to the optimizer. Returns ------- MLEResults See Also -------- statsmodels.base.model.LikelihoodModel.fit statsmodels.tsa.statespace.mlemodel.MLEResults emr transformedcov_typecov_kwdsr toleranceem_initialization mstep_method full_outputrr low_memoryllf_decrease_actionllf_decrease_tolerancerrincludes_fixedrrrrrrcallbackr optim_scoreoptim_complex_step optim_hessianflagsrrO)rrr8)r9rrrrrrrrrrrrrrrrrrrrrrrs r:r8zDynamicFactorMQ.fitsX T>4;;I){!Hg#7H){+ $7'=IBHI I7; 1) 17B 1- 18@ 1" 1+1 1;B 1( 1/3 1 " 1 2? 1 ( 1$6 1, 149 1& 1 1r<c>|jr td| r td||j}d}nt j |d}|s|j |}t| dgd } t|}|j}g}|g}d}|jjg}d }d }d }||kry|sv|dks||kDrk|j|d || }|d jj}|s|j|dg}|d}|jj}t j |j"}|j$d k(r|j&s|j(d }|j*dk(r?||d|df}t j,tj.j1|dks|j3|dd|j5dt7|j8n|j(D]}t7||d}|j:|j*}|dk(s3||d|df}t j,tj.j1|dkrv|j3|dd|j5dt7|j8|j&r\|j=d} t?|dj@|djBD]{}!|j:|!fj*}|dk(s#t jD||!|!fdkrA|j3|!d| |!|dj@z }"|j5d|"}|j$d kDrt7||d}|j:|j*}|dk(rh||d|df}t j,tj.j1|dks&|j3|dd|j5dtG|d kDrbtId|dzd|d|jK|d ||||||| | || | | |}#|jjM|jO|#S|d kDxr ||d z | k}$| dk(r |$rtId|dzd|d |dz}d}n@| d!k(r|$rtId|dzd"|j5||j5|d|rTtjP|j"d#|d jRd$|d jTd$%}|j5||d kDrTd&t jD|d |d'z zt jD|d t jD|d'zz }ntjV}|r|d k(rtYd(|d d)n)|r'|dz|zd k(rtYd*|dzd+|d d)d,|d)|dz }||kr|s|dkrd||kDrk||k(xr||kD}%|%rtId-|d.|d d)d,|d)d/|d)d0 |r_|rtYd1|d+|d d)d,|d)d/|d)d0 n>|%rtYd-|d.|d d)d,|d)d/|d)d0 ntYd2|d+|d d)d,|d)d3|d)| r|d }&|&S|r'|jj}'||j_ |j[|d d||4}&|r'|j_ | ra|j5|&j\t_d8it j |t j |||d5}(t_d8id6||d7})nd}(d})|(|&j`_1|)|&j`_2|&S)9a Fits the model by maximum likelihood via the EM algorithm. Parameters ---------- start_params : array_like, optional Initial guess of the solution for the loglikelihood maximization. The default is to use `DynamicFactorMQ.start_params`. transformed : bool, optional Whether or not `start_params` is already transformed. Default is True. cov_type : str, optional The `cov_type` keyword governs the method for calculating the covariance matrix of parameter estimates. Can be one of: - 'opg' for the outer product of gradient estimator - 'oim' for the observed information matrix estimator, calculated using the method of Harvey (1989) - 'approx' for the observed information matrix estimator, calculated using a numerical approximation of the Hessian matrix. - 'robust' for an approximate (quasi-maximum likelihood) covariance matrix that may be valid even in the presence of some misspecifications. Intermediate calculations use the 'oim' method. - 'robust_approx' is the same as 'robust' except that the intermediate calculations use the 'approx' method. - 'none' for no covariance matrix calculation. Default is 'none', since computing this matrix can be very slow when there are a large number of parameters. cov_kwds : dict or None, optional A dictionary of arguments affecting covariance matrix computation. **opg, oim, approx, robust, robust_approx** - 'approx_complex_step' : bool, optional - If True, numerical approximations are computed using complex-step methods. If False, numerical approximations are computed using finite difference methods. Default is True. - 'approx_centered' : bool, optional - If True, numerical approximations computed using finite difference methods use a centered approximation. Default is False. maxiter : int, optional The maximum number of EM iterations to perform. tolerance : float, optional Parameter governing convergence of the EM algorithm. The `tolerance` is the minimum relative increase in the likelihood for which convergence will be declared. A smaller value for the `tolerance` will typically yield more precise parameter estimates, but will typically require more EM iterations. Default is 1e-6. disp : int or bool, optional Controls printing of EM iteration progress. If an integer, progress is printed at every `disp` iterations. A value of True is interpreted as the value of 1. Default is False (nothing will be printed). em_initialization : bool, optional Whether or not to also update the Kalman filter initialization using the EM algorithm. Default is True. mstep_method : {None, 'missing', 'nonmissing'}, optional The EM algorithm maximization step. If there are no NaN values in the dataset, this can be set to "nonmissing" (which is slightly faster) or "missing", otherwise it must be "missing". Default is "nonmissing" if there are no NaN values or "missing" if there are. full_output : bool, optional Set to True to have all available output from EM iterations in the Results object's mle_retvals attribute. return_params : bool, optional Whether or not to return only the array of maximizing parameters. Default is False. low_memory : bool, optional This option cannot be used with the EM algorithm and will raise an error if set to True. Default is False. llf_decrease_action : {'ignore', 'warn', 'revert'}, optional Action to take if the log-likelihood decreases in an EM iteration. 'ignore' continues the iterations, 'warn' issues a warning but continues the iterations, while 'revert' ends the iterations and returns the result from the last good iteration. Default is 'warn'. llf_decrease_tolerance : float, optional Minimum size of the log-likelihood decrease required to trigger a warning or to end the EM iterations. Setting this value slightly larger than zero allows small decreases in the log-likelihood that may be caused by numerical issues. If set to zero, then any decrease will trigger the `llf_decrease_action`. Default is 1e-4. Returns ------- DynamicFactorMQResults See Also -------- statsmodels.tsa.statespace.mlemodel.MLEModel.fit statsmodels.tsa.statespace.mlemodel.MLEResults zFCannot fit using the EM algorithm while holding some parameters fixed.z?Cannot fit using the EM algorithm when using low_memory option.NTrS)ndminr)ignorerrevertoptionsrFr)rArrr>r)gA?diffusezfactor block: ryrz*idiosyncratic AR(1) for monthly variable: rz8idiosyncratic AR(1) for the block of quarterly variablesz0Non-stationary parameters found at EM iteration zu, which is not compatible with stationary initialization. Initialization was switched to diffuse for the following: z, and fitting was restarted.rrz)Log-likelihood decreased at EM iteration z-. Reverting to the results from EM iteration z4 (prior to the decrease) and returning the solution.rz&, which can indicate numerical issues.known.r)constantstationary_covrzEM start iterations, llf=z.5gz EM iteration z, llf=z, convergence criterion=z)EM reached maximum number of iterations (z&), without achieving convergence: llf=z (while specified tolerance was )zEM terminated at iteration zEM converged at iteration z < tolerance=)rrr)rllfiterinitsr)rrrrO)3_has_fixed_paramsNotImplementedErrorrjrr>rrrrrrr _em_iterationllf_obsrtrr@r8r5rbrinitialization_typeallreigvalsrnrrhr.blocksrFrkrr@absr/rrrrr?smoothed_statesmoothed_state_covrprintsmoothrr_results mle_retvals mle_settings)*r9rrrrrrrrrrrrrrrrrrArrdelta terminateoutnew_llf switch_initrArrYTbb init_typerrr[results llf_decrease not_convergedresult base_init em_retvals em_settingss* r:rzDynamicFactorMQ.fit_em sD  ! !%'HI I 34 4  ,,LK88L:L00>L) !6024y GG(()  'k)Q59;L$$VBZd2>%@C!fnn((*G% CF# &xx..yy/>>Q&t/E/EOOA.E//<?uY/y1AAB!vvbii&7&7&;y&IJ HHU9%5yA'..!$$)%*<*<$=#>!@A "# E!$uY'7"89$(KKN$F$F $ 4!"5#3U95E#E!FB#%66"))*;*;B*?9*M#N $y)99 E + 2 2%((-e.@.@(A'B%D!E E))"&"7"7$"7"GK"1[>#7#7;9L9LM:$(KK$5$I$I $ 4#%66!AqD'?i#@ $I 6'21q~7K7K3K'L + 2 2%226%9!: :~~)!$q~"67$(KKN$F$F $ 4!"1[>1[>#A!BB#%66"))*;*;B*?9*M#N $; C + 2 2%D!E{#a'UG$)M)E GH #kk%+BZ[!)H '9*;%1{!#-,?/E* GGHH''(D(D(FG"N AG7SW,1G0GG #h.<@QHDDE3KLMQ &&0\DQUGLAAB 7# c!f%$)88 w!$Q!6!6v!>'*1v'@'@'HJDLL&q5B#b'(9!:: ffSWos2w?AEFFEAF5c"gc]CDA~!3M!a%s2wsmD449#;@A FAs'k)Q59;Lxg;%)*;    9! HH33 *.'[[*2X!GF *3' 6::&"70@-/XXc].//4&67 $<4=29';< " " *4FOO '+6FOO ( r<cX|j||}|j|||}||fS)z EM iteration.)rA)r)_em_expectation_step_em_maximization_step)r9params0rArresparams1s r:rzDynamicFactorMQ._em_iteration sD''d';,,S':F-HG|r<cr|j||'|jj}||j_|jjtt zt zd}tj|jjjd|_ ||j_|S)zEM expectation step.F) update_filterT)r$) rrrr rrr r>r_kalman_filter loglikelihoodr)r9r"rArr#s r:r z$DynamicFactorMQ._em_expectation_step s G  //I&*DHH #hhoo / /2H H!hh HH # # 1 1>   &/DHH # r<c  |j}|jjd}|jj ddd}|j j ddd}|j tj||j dddz}|ddtj|dd|ddj dddz} tj|j} || rdnd}|j}|dk(r | r td |dk(r |j} n |dk(r |j} ntd |z| ||||j \} } g}g}|j D]}|dd|d |d fj#d }| dd|d|d fj#d }|dd|d|dfj#d }|j$ddz } t't)||jj}|||jzz |z }||j1j3z }|tj,j5|tj6|j3z }|jr|d}|dd||fj#d j9}| dd||fj#d j9}|dd||fj#d j9}|j$ddz }||z }|||zz |z }np|d}|dd||fj#d }tj:| j9|j<d|j9|j$dz f}tj>|}g}tA|jBD];}|jjD|} |d| }!|| ||!fj3z }=|||jFd<|||jFd<|||jFd<|jr||jFd<|||jFd<|S#t*$r8tj,j/||jj}YwxYw)zEM maximization step.rrrrSNrr nonmissingzMCannot use EM algorithm option `mstep_method="nonmissing"` with missing data.z'Invalid maximization step method: "%s".) compute_Hr*rTr,rrrrrrbr)$rr rAr rsmoothed_state_autocovr$r>matmulrnmissinglowerrj_em_maximization_obs_nonmissing_em_maximization_obs_missingrbrrtrsrrr rsolverrrrdiagonalr_r zeros_likerkrarr )"r9r#r"rracov_aacov_aEaaEaa1 has_missingfuncLambdaHrrrrBCr[f_Af_QrAdBdCdrsigma2r$rrrrs" r:r!z%DynamicFactorMQ._em_maximization_step s GG     +&&00Aq9++55aA>jjlRYYq!++aA*>??cr{RYYquaf.>.>q!Q.GHHffS\\*  (39L#))+ < 'KOP P < '77D Y &44DF+,- -c1T5K5K1KM    MACRC<!L/9:>>A>FAQ,<89==1=EAAB,<89==1=EA99Q=R>c@|j}|j}|j}tj|j |f|}t |j D]} |jdd| | dzf} |jj| } |d| } tjddftj| | z} || jd}| j|dd| dfz}|jr3|j| z}|dz}||dd||| fjdz} tt||jj|| | f<|r|dj'}||ddd|f<|jj|dz|jz}|jj|jz}|| z|jz ||jdz|jzz|j(z }||fStj|j |j f|tj*z}||fS#t $r=tj"j%||jj|| | f<Y?wxYw) z@EM maximization step, observation equation without missing data.r^NrSr,rrTrr)rr_rr>rrarkrrrEix_rtrArbrrr rr2r$r[r)r9r#r9r6r+rr_rr=rrrrrrr?rrZBLr@r>s r:r0z/DynamicFactorMQ._em_maximization_obs_nonmissing5 s_ GG    4<<+59t||$ AA 1aAg:&A77,,Q/D,-I%%(rvvi;;BB  #Aa9a((A%%((1,AgSCGY./333;; A'0A'D'F'Fq)|$ A6 X##%AAa!eH& )ACC/B tzz)AbS244!cgg1go"5";;tyyHAqy$,, 5UCbffLAqy+ A(*yyq!##'>'@'@q)|$ As1.IAJJc\ |j}|j}|j}tj|j |f|}d|j jz } | jt} t|jD]} |jj| } |d| } | dd| f}|j|| | dzf}tj|| | }||jd}|j|tj|| dz}|j r5|j| z}|dz}|||||fd| fjdz} t#t%||jj|| | f< |j,dkDrtj.gd dddf}t|j|j D]} |jj| } |d dd| fj1j3} |j5| \}}| |jz }| dd| f}|j|| | dzf}tj|| | }||jd}|j|tj|| dz}|j ratj6ddftj|d || z}||}||||jdzjdz} t%|}t#||jjd}t#||j}t%||z}t#||}|||z|zz }||| | f<|r|d j;}"||"ddd|j<df<tj>|j}#|#j|#z}$| d } d| z }%| |"z}&|&jAddd}'|#d |jAdddz|'z}(|})|(jAddd}*|%jAddd}+|$|( |*z |&|)z|'zz|%|dz|+zzjdz|jBz },||,fStj|j |j f|tjDz},||,fS#t&$r=tj(j+||jj|| | f<YwxYw#t&$rutj(j9|} || zd}tj(j9|| z|jz}!|| |jz|!z|z|zz }YwxYw)z=EM maximization step, observation equation with missing data.r^rSr,NrrTr.rrrrrrr)#rr_rr>rrarrAr*rrkr`rrrHrtrbrrr rr2r5rrrrrEinvr$rs nan_to_numrr[r)-r9r#r9r6r+rr_rr=rr1rrrrytrAiBirrrrriQBiQEepsf L_and_lower unrestrictedAiiRTRAiiRTiR restrictedAiiRARirIrr@IWWLWLTrJrBLTIWTr>s- r:r1z,DynamicFactorMQ._em_maximization_obs_missingg sw GG    4<<+59  xx~t~~& CA77,,Q/D,-IQT AAqQwJ'B9i0BR!$B"&&I./77B%%((1,Agc!SW*oc9n599q9AA C'0B'F'H'Hq)|$! C2 >>A ((?3AtG>4<<8) 2ww003/0D9??AHHJ //21'AJZZ1QU7 +VVAy)4W[[a[(ddQrvva34V<<))%%(rvva.?.CY'OOBGEK%(,,A,*>>CCCKKCG",R.K#,[#%%#@#B#B1#EL%k1337E",QY"7K(a8H!-0@<0O!OJ(2q)|$S) 2b X##%A%+Aa!&,,q/!! " djj)AaA) AQBQB,,q!Q'C9 Aq! 44s:BA,,q!Q'C,,q!Q'CrcCi"q&3,.4 ?*S0125#1#+>AEKA qy$,, 5UCbffLAqye C(*yyr244'@'B'Bq)|$ CX#G))--+C$'#Iq>L99==S1337D"."%)d"2Q"6"E#FJ Gs'&.S$/A)T-$AT*)T*-A:V+*V+c 6t ||f|||||||| d| S)a Kalman smoothing. Parameters ---------- params : array_like Array of parameters at which to evaluate the loglikelihood function. transformed : bool, optional Whether or not `params` is already transformed. Default is True. return_ssm : bool,optional Whether or not to return only the state space output or a full results object. Default is to return a full results object. cov_type : str, optional See `MLEResults.fit` for a description of covariance matrix types for results object. Default is None. cov_kwds : dict or None, optional See `MLEResults.get_robustcov_results` for a description required keywords for alternative covariance estimators **kwargs Additional keyword arguments to pass to the Kalman filter. See `KalmanFilter.filter` for more details. rr complex_steprr return_ssm results_classresults_wrapper_class)rr ) r9rrrrbrrrcrdrerrs r:r zDynamicFactorMQ.smooth sA6w~ C +N%8!"7 C 1` and `repetitions` is not None, then the output will be a Pandas DataFrame that has a MultiIndex for the columns, with the first level containing the names of the `endog` variables and the second level containing the repetition number. ) measurement_shocks state_shocks initial_stateanchor repetitionsr extend_model extend_kwargsrrrSrrrUlevelr)rsimulaterrfrr!rsr/rrrmultiplyaddr>r)r9r nsimulationsrirjrkrlrmrrnrorroriginal_scalersim use_pandasrsrr|rs r:rrzDynamicFactorMQ.simulate! s`g LN5G%]{%]#N N GM N   #DIIz:JIIEu:?//..q1C++003D)d*C, %Z1_<<aq<I"s4#3#3!1sE" u:?/$2B2BBC Z1_/$2B2BBC --8CC==)9)9:9ED)d*C r<c  t||f|||||||| | | d | }|jr| rt|jt }|j }|rZt|dk(r||jjdz}|St|dk(r|j|jdd}|St|dk(r||jz}|St|dk(r||jz}|S)a[ Impulse response function. Parameters ---------- params : array_like Array of model parameters. steps : int, optional The number of steps for which impulse responses are calculated. Default is 1. Note that for time-invariant models, the initial impulse is not counted as a step, so if `steps=1`, the output will have 2 entries. impulse : int or array_like If an integer, the state innovation to pulse; must be between 0 and `k_posdef-1`. Alternatively, a custom impulse vector may be provided; must be shaped `k_posdef x 1`. orthogonalized : bool, optional Whether or not to perform impulse using orthogonalized innovations. Note that this will also affect custum `impulse` vectors. Default is False. cumulative : bool, optional Whether or not to return cumulative impulse responses. Default is False. anchor : int, str, or datetime, optional Time point within the sample for the state innovation impulse. Type depends on the index of the given `endog` in the model. Two special cases are the strings 'start' and 'end', which refer to setting the impulse at the first and last points of the sample, respectively. Integer values can run from 0 to `nobs - 1`, or can be negative to apply negative indexing. Finally, if a date/time index was provided to the model, then this argument can be a date string to parse or a datetime type. Default is 'start'. exog : array_like, optional New observations of exogenous regressors for our-of-sample periods, if applicable. transformed : bool, optional Whether or not `params` is already transformed. Default is True. includes_fixed : bool, optional If parameters were previously fixed with the `fix_params` method, this argument describes whether or not `params` also includes the fixed parameters, in addition to the free parameters. Default is False. original_scale : bool, optional If the model specification standardized the data, whether or not to return impulse responses in the original scale of the data (i.e. before it was standardized by the model). Default is True. **kwargs If the model has time-varying design or transition matrices and the combination of `anchor` and `steps` implies creating impulse responses for the out-of-sample period, then these matrices must have updated values provided for the out-of-sample steps. For example, if `design` is a time-varying component, `nobs` is 10, `anchor=1`, and `steps` is 15, a (`k_endog` x `k_states` x 7) matrix must be provided with the new design matrix values. Returns ------- impulse_responses : ndarray Responses for each endogenous variable due to the impulse given by the `impulse` argument. For a time-invariant model, the impulse responses are given for `steps + 1` elements (this gives the "initial impulse" followed by `steps` responses for the important cases of VAR and SARIMAX models), while for time-varying models the impulse responses are only given for `steps` elements (to avoid having to unexpectedly provide updated time-varying matrices). ) stepsimpulseorthogonalized cumulativerlrrnrorrrSrrrp) rimpulse_responsesrrfrr!rsr/rrrs)r9rrzr{r|r}rlrrnrorrrvrirfsrxrsrs r:r~z!DynamicFactorMQ.impulse_responses sVw( 5)j<'[) 5 .4 5   #DIIz:JJJEu:?$//"6"6q"99D Z1_==q=JD u:?$//1D  Z1_$//1D r<) NrSrSNTTNFF)NNF)NNr)NTFnoneNrư>TNrSFNFNNNNFrr)NTrNrrFTNTFFrr)F)TFFrNFNN) TFFrNFNNF) NNNNNNNNTFT) rSrFFNNNNTFT)#rJrKrLrMr; classmethodrr6rNr<rrFrLrjrrrur|rrrrrrr8rrr r!r0r1r rgrrr~ __classcell__rs@r:rrsG RNO?CAF#S2jc&c&JBF%*BMM 2  jX#.''RIIV0d/bM+^KKT,lGLADAEDIEI>F#' _1BDJ@EFJ/4DH xt (Un0dqf?D=A/3%)CB?D=A/36;$CLAE?C;?FK $ wr:;IMFJ;@)- hhr<rceZdZdZdfd ZedZ ddZedZ ddZ dfd Z dfd Z dfd Z dfd Z dfd Z dd Z dfd ZxZS)r9z5 Results from fitting a dynamic factor model c ,t|||||fi|yr)rr;)r9r\rfilter_resultsrrrs r:r;zDynamicFactorMQResults.__init__ s!  6>8 ?7= ?r<cd}|jjdkDrI|jjj}t j |jj |j}t|jjjdd|f|jjjtj|ddf|fdd}|j,|jjjdd|f|_|j ?|jj"jtj|ddf|f|_|S)a Estimates of unobserved factors. Returns ------- out : Bunch Has the following attributes shown in Notes. Notes ----- The output is a bunch of the following format: - `filtered`: a time series array with the filtered estimate of the component - `filtered_cov`: a time series array with the filtered estimate of the variance/covariance of the component - `smoothed`: a time series array with the smoothed estimate of the component - `smoothed_cov`: a time series array with the smoothed estimate of the variance/covariance of the component - `offset`: an integer giving the offset in the state vector where this component begins Nr)filtered filtered_covsmoothed smoothed_cov)r\r0rr,r>rrurrstatesrr2rrEr rr r)r9rrrs r:r)zDynamicFactorMQResults.factors s2 ::  ! #::==++D$**001$7>>@B--11!R%8![[5599"%%A,:JKD2C"".#{{33772> &&2KK,,00r1ur1AB  r<c ddlm}t|dgd}||jdnd}|jj }|jj }|jjj}|jj}|jj}|d k(rtj||f} t|D]} ||j|jdd| f} t|D]m} |j| | frE|j j"| } t%| | d j'j(| | | f<Ytj*| | | f<ot-j.| || } | S|d k(rtj|f} ||j|} t|D]} |j j"| } tj0d|j| fj3}| j4dd|f}t%| |d j'j(| | <t-j6| |} | S|dk(rWtj||f} ||j|} t|D]} |j j"| } t|D]} |jjjj| | frtj0d|j| d| dzfdg|| z dz zf}| j4dd|j9t:j3f}t%| |d j'j(| | | f<tj*| | | f<t-j.| || }  S)a Get coefficients of determination (R-squared) for variables / factors. Parameters ---------- method : {'individual', 'joint', 'cumulative'}, optional The type of R-squared values to generate. "individual" plots the R-squared of each variable on each factor; "joint" plots the R-squared of each variable on each factor that it loads on; "cumulative" plots the successive R-squared values as each additional factor is added to the regression, for each variable. Default is 'individual'. which: {None, 'filtered', 'smoothed'}, optional Whether to compute R-squared values based on filtered or smoothed estimates of the factors. Default is 'smoothed' if smoothed results are available and 'filtered' otherwise. Returns ------- rsquared : pd.DataFrame or pd.Series The R-squared values from regressions of observed variables on one or more of the factors. If method='individual' or method='cumulative', this will be a Pandas DataFrame with observed variables as the index and factors as the columns . If method='joint', will be a Pandas Series with observed variables as the index. See Also -------- plot_coefficients_of_determination coefficients_of_determination r) add_constantr individualjointr}rNrrrrrrrTrr}rSF)statsmodels.toolsrrr r\rar0rr3rr.r>rrkr)rrrr r8rsquaredrr}rr4rr2rr*r)r9rwhichrrar0ef_maprr. coefficientsrrrrrrRs r:!get_coefficients_of_determinationz8DynamicFactorMQResults.get_coefficients_of_determination5 sD 3VX8FG ="&"5"5"=J:E**$$JJ(( //jj,, zz.. \ !88Wi$89L9% 4#DLL$7$<$L<9w 88WJ/L U 34D7^ B++11!4UU4Q/0779HHQUOq&1557@@Q  B 99\EL&%| #88Wi$89L U 34D7^ 4++11!4y)4Azz}}55::1a4@UU4QQY)?$)7i!ma.?#@$AB HHQ $(>(>(@%@Aq&9==?HH%QT*.0VV QT*4 4<< K0<>Lr<c&|jdS)a Individual coefficients of determination (:math:`R^2`). Coefficients of determination (:math:`R^2`) from regressions of endogenous variables on individual estimated factors. Returns ------- coefficients_of_determination : ndarray A `k_endog` x `k_factors` array, where `coefficients_of_determination[i, j]` represents the :math:`R^2` value from a regression of factor `j` and a constant on endogenous variable `i`. Notes ----- Although it can be difficult to interpret the estimated factor loadings and factors, it is often helpful to use the coefficients of determination from univariate regressions to assess the importance of each factor in explaining the variation in each endogenous variable. In models with many variables and factors, this can sometimes lend interpretation to the factors (for example sometimes one factor will load primarily on real variables and another on nominal variables). See Also -------- get_coefficients_of_determination plot_coefficients_of_determination r)r)rr;s r:coefficients_of_determinationz4DynamicFactorMQResults.coefficients_of_determination s@55\5JJr<cddlm}m}||||}t|dgd}||jj dk}|j ||}|dvrd } |jjD]\} } |j|jjd | } | jd | j| d | j| d | t|jks|s| j j#g| d z } |S|dk(rg|jd d d } | jd | jdd |j| d |s| j j#g|S)a Plot coefficients of determination (R-squared) for variables / factors. Parameters ---------- method : {'individual', 'joint', 'cumulative'}, optional The type of R-squared values to generate. "individual" plots the R-squared of each variable on each factor; "joint" plots the R-squared of each variable on each factor that it loads on; "cumulative" plots the successive R-squared values as each additional factor is added to the regression, for each variable. Default is 'individual'. which: {None, 'filtered', 'smoothed'}, optional Whether to compute R-squared values based on filtered or smoothed estimates of the factors. Default is 'smoothed' if smoothed results are available and 'filtered' otherwise. endog_labels : bool, optional Whether or not to label the endogenous variables along the x-axis of the plots. Default is to include labels if there are 5 or fewer endogenous variables. fig : Figure, optional If given, subplots are created in this figure instead of in a new figure. Note that the grid will be created in the provided figure using `fig.add_subplot()`. figsize : tuple, optional If a figure is created, this argument allows specifying a size. The tuple is (width, height). Notes ----- The endogenous variables are arranged along the x-axis according to their position in the model's `endog` array. See Also -------- get_coefficients_of_determination r) _import_mplcreate_mpl_figrrrr()rr)rr}rS)rrSz$R^2$)rQylabelbar)axkindrz($R^2$ - regression on all loaded factors)statsmodels.graphics.utilsrrrr\rarrAiterrows add_subplotr0set_ylimrnplotr/rxaxisset_ticklabels) r9rr endog_labelsfigfigsizerrrplot_idxrcoeffsrs r:"plot_coefficients_of_determinationz9DynamicFactorMQResults.plot_coefficients_of_determination sP K S'*VX8FG  ::--2L99@E:G 1 1H'/zz':':'< # V__TZZ%9%91hG F# }h? r .c("2"233<HH++B/A  ( w Aq)B KK  FFD"  $ MMReM ,''+ r<c t|d|||||||| | d | } |jjr.|r+| j} | j j \}}tj|jj}tj|jj}|jjdkDr|dddf}|dddf}| jj|z|z| j_ |dk(r$| jxj|dzzc_| S|| jjz|jz| j_| S)a In-sample prediction and out-of-sample forecasting. Parameters ---------- start : int, str, or datetime, optional Zero-indexed observation number at which to start forecasting, i.e., the first forecast is start. Can also be a date string to parse or a datetime type. Default is the the zeroth observation. end : int, str, or datetime, optional Zero-indexed observation number at which to end forecasting, i.e., the last forecast is end. Can also be a date string to parse or a datetime type. However, if the dates index does not have a fixed frequency, end must be an integer index if you want out of sample prediction. Default is the last observation in the sample. dynamic : bool, int, str, or datetime, optional Integer offset relative to `start` at which to begin dynamic prediction. Can also be an absolute date string to parse or a datetime type (these are not interpreted as offsets). Prior to this observation, true endogenous values will be used for prediction; starting with this observation and continuing through the end of prediction, forecasted endogenous values will be used instead. information_set : str, optional The information set to condition each prediction on. Default is "predicted", which computes predictions of period t values conditional on observed data through period t-1; these are one-step-ahead predictions, and correspond with the typical `fittedvalues` results attribute. Alternatives are "filtered", which computes predictions of period t values conditional on observed data through period t, and "smoothed", which computes predictions of period t values conditional on the entire dataset (including also future observations t+1, t+2, ...). signal_only : bool, optional Whether to compute forecasts of only the "signal" component of the observation equation. Default is False. For example, the observation equation of a time-invariant model is :math:`y_t = d + Z \alpha_t + \varepsilon_t`, and the "signal" component is then :math:`Z \alpha_t`. If this argument is set to True, then forecasts of the "signal" :math:`Z \alpha_t` will be returned. Otherwise, the default is for forecasts of :math:`y_t` to be returned. original_scale : bool, optional If the model specification standardized the data, whether or not to return predictions in the original scale of the data (i.e. before it was standardized by the model). Default is True. **kwargs Additional arguments may required for forecasting beyond the end of the sample. See `FilterResults.predict` for more details. Returns ------- forecast : ndarray Array of out of in-sample predictions and / or out-of-sample forecasts. An (npredict x k_endog) array. ) renddynamicinformation_set signal_onlyr\rrnrorSNrrO)rget_predictionr\rprediction_resultsrrsr>rrrrar _predicted_mean_var_pred_meanrA)r9rrrrrrvr\rrnrorr#rrarr|rrs r:rz%DynamicFactorMQResults.get_predictionsL|g$L5c75D1<+0t2>3@ L EK L :: ! !n!$!7!7 +1177JGQ88DJJ223D((4::001Czz!!A%D!G}$'l ,,s2T9 LL (!| ++sAv5+  #,,555= + r<c | dk(r@tj|jj|jjz } t ||f|||||||| | | | | d |}| s|jjr|r|jj}|jj}|j|z|_ |j|z|_ |j|j|z|_ |j|j|z|_ |j|j|z|_dD]h}t||}d}t!|dr |j"}|j%|d}|dvr|j'|d}|||_t)|||j|j*j-|d d j%|dd |_|j.j-|d d j%|dd |_|S) u Compute impacts from updated data (news and revisions). Parameters ---------- comparison : array_like or MLEResults An updated dataset with updated and/or revised data from which the news can be computed, or an updated or previous results object to use in computing the news. impact_date : int, str, or datetime, optional A single specific period of impacts from news and revisions to compute. Can also be a date string to parse or a datetime type. This argument cannot be used in combination with `start`, `end`, or `periods`. Default is the first out-of-sample observation. impacted_variable : str, list, array, or slice, optional Observation variable label or slice of labels specifying that only specific impacted variables should be shown in the News output. The impacted variable(s) describe the variables that were *affected* by the news. If you do not know the labels for the variables, check the `endog_names` attribute of the model instance. start : int, str, or datetime, optional The first period of impacts from news and revisions to compute. Can also be a date string to parse or a datetime type. Default is the first out-of-sample observation. end : int, str, or datetime, optional The last period of impacts from news and revisions to compute. Can also be a date string to parse or a datetime type. Default is the first out-of-sample observation. periods : int, optional The number of periods of impacts from news and revisions to compute. exog : array_like, optional Array of exogenous regressors for the out-of-sample period, if applicable. comparison_type : {None, 'previous', 'updated'} This denotes whether the `comparison` argument represents a *previous* results object or dataset or an *updated* results object or dataset. If not specified, then an attempt is made to determine the comparison type. state_index : array_like or "common", optional An optional index specifying a subset of states to use when constructing the impacts of revisions and news. For example, if `state_index=[0, 1]` is passed, then only the impacts to the observed variables arising from the impacts to the first two states will be returned. If the string "common" is passed and the model includes idiosyncratic AR(1) components, news will only be computed based on the common states. Default is to use all states. return_raw : bool, optional Whether or not to return only the specific output or a full results object. Default is to return a full results object. tolerance : float, optional The numerical threshold for determining zero impact. Default is that any impact less than 1e-10 is assumed to be zero. endog_quarterly : array_like, optional New observations of quarterly variables, if `comparison` was provided as an updated monthly dataset. If this argument is provided, it must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. References ---------- .. [1] Bańbura, Marta, and Michele Modugno. "Maximum likelihood estimation of factor models on datasets with arbitrary pattern of missing data." Journal of Applied Econometrics 29, no. 1 (2014): 133-160. .. [2] Bańbura, Marta, Domenico Giannone, and Lucrezia Reichlin. "Nowcasting." The Oxford Handbook of Economic Forecasting. July 8, 2011. .. [3] Bańbura, Marta, Domenico Giannone, Michele Modugno, and Lucrezia Reichlin. "Now-casting and the real-time data flow." In Handbook of economic forecasting, vol. 2, pp. 195-237. Elsevier, 2013. common) impact_dateimpacted_variablerrperiodsrcomparison_typerevisions_details_start state_index return_rawrrN) prev_impacted_forecastsnews revisionsupdate_realizedupdate_forecastsrevised revised_prevpost_impacted_forecasts revisions_all revised_allrevised_prev_allr[rS)rq)rrrrp)r>r@r\r8rarrrrr total_impactsupdate_impactsrevision_impactsrevision_detailed_impactsrevision_grouped_impactsrr#r[rsrtsetattrweightsdividerevision_weights)r9 comparisonrrrrrrrrrrrrrvr news_resultsrrr[dta orig_namers r:rzDynamicFactorMQResults.news^s7^ ( " $**-- 0B0BBC w| L$//u#$$;#  L EK L djj44//J --I **Y6  &++i7  ',,8 11I=-55A ::YF644@ 99IE5M 1lD1 ! 3' #Ill9Al644''*A'6C((CH dC0- 1F$$++IAQ+G%-XiaqX%I  --#VIAQV?%XiaqXA  ) r<c$|jjr4|r2|jd}|jj|jd<t |||\}}}}|jjr|r|jd<|dk(r~|jjrh|rf|jj } |j | dd}|j | dd}|j | dd}|j | dd}||||fS)a Decompose smoothed output into contributions from observations Parameters ---------- decomposition_of : {"smoothed_state", "smoothed_signal"} The object to perform a decomposition of. If it is set to "smoothed_state", then the elements of the smoothed state vector are decomposed into the contributions of each observation. If it is set to "smoothed_signal", then the predictions of the observation vector based on the smoothed state vector are decomposed. Default is "smoothed_state". state_index : array_like, optional An optional index specifying a subset of states to use when constructing the decomposition of the "smoothed_signal". For example, if `state_index=[0, 1]` is passed, then only the contributions of observed variables to the smoothed signal arising from the first two states will be returned. Note that if not all states are used, the contributions will not sum to the smoothed signal. Default is to use all states. original_scale : bool, optional If the model specification standardized the data, whether or not to return simulations in the original scale of the data (i.e. before it was standardized by the model). Default is True. Returns ------- data_contributions : pd.DataFrame Contributions of observations to the decomposed object. If the smoothed state is being decomposed, then `data_contributions` is shaped `(k_states x nobs, k_endog x nobs)` with a `pd.MultiIndex` index corresponding to `state_to x date_to` and `pd.MultiIndex` columns corresponding to `variable_from x date_from`. If the smoothed signal is being decomposed, then `data_contributions` is shaped `(k_endog x nobs, k_endog x nobs)` with `pd.MultiIndex`-es corresponding to `variable_to x date_to` and `variable_from x date_from`. obs_intercept_contributions : pd.DataFrame Contributions of the observation intercept to the decomposed object. If the smoothed state is being decomposed, then `obs_intercept_contributions` is shaped `(k_states x nobs, k_endog x nobs)` with a `pd.MultiIndex` index corresponding to `state_to x date_to` and `pd.MultiIndex` columns corresponding to `obs_intercept_from x date_from`. If the smoothed signal is being decomposed, then `obs_intercept_contributions` is shaped `(k_endog x nobs, k_endog x nobs)` with `pd.MultiIndex`-es corresponding to `variable_to x date_to` and `obs_intercept_from x date_from`. state_intercept_contributions : pd.DataFrame Contributions of the state intercept to the decomposed object. If the smoothed state is being decomposed, then `state_intercept_contributions` is shaped `(k_states x nobs, k_states x nobs)` with a `pd.MultiIndex` index corresponding to `state_to x date_to` and `pd.MultiIndex` columns corresponding to `state_intercept_from x date_from`. If the smoothed signal is being decomposed, then `state_intercept_contributions` is shaped `(k_endog x nobs, k_states x nobs)` with `pd.MultiIndex`-es corresponding to `variable_to x date_to` and `state_intercept_from x date_from`. prior_contributions : pd.DataFrame Contributions of the prior to the decomposed object. If the smoothed state is being decomposed, then `prior_contributions` is shaped `(nobs x k_states, k_states)`, with a `pd.MultiIndex` index corresponding to `state_to x date_to` and columns corresponding to elements of the prior mean (aka "initial state"). If the smoothed signal is being decomposed, then `prior_contributions` is shaped `(nobs x k_endog, k_states)`, with a `pd.MultiIndex` index corresponding to `variable_to x date_to` and columns corresponding to elements of the prior mean. Notes ----- Denote the smoothed state at time :math:`t` by :math:`\alpha_t`. Then the smoothed signal is :math:`Z_t \alpha_t`, where :math:`Z_t` is the design matrix operative at time :math:`t`. obs_intercept)decomposition_ofrsmoothed_signalrrp)r\rrrget_smoothed_decompositionrrs) r9rrrvcache_obs_interceptdata_contributionsobs_intercept_contributionsstate_intercept_contributionsprior_contributionsrrs r:rz1DynamicFactorMQResults.get_smoothed_decompositionsCj :: ! !n"&**_"= *.***@*@DJJ ' G .!1{ / L = 8 &(; :: ! !n*=DJJ '  1 1JJ**~ --I#++IAQ+G ,44AQ50 (.66AQ70 *$,,YQa,H #$?-/BD Dr<c 6tj||\}}t|jdk(r|jdnd} ||jj k7s| |jj k7r td||d<t |$|f||||d|S)a Recreate the results object with new data appended to original data. Creates a new result object applied to a dataset that is created by appending new data to the end of the model's original data. The new results can then be used for analysis or forecasting. Parameters ---------- endog : array_like New observations from the modeled time-series process. endog_quarterly : array_like, optional New observations of quarterly variables. If provided, must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. refit : bool, optional Whether to re-fit the parameters, based on the combined dataset. Default is False (so parameters from the current results object are used to create the new results object). fit_kwargs : dict, optional Keyword arguments to pass to `fit` (if `refit=True`) or `filter` / `smooth`. copy_initialization : bool, optional Whether or not to copy the initialization from the current results set to the new model. Default is True. retain_standardization : bool, optional Whether or not to use the mean and standard deviations that were used to standardize the data in the current model in the new model. Default is True. **kwargs Keyword arguments may be used to modify model specification arguments when created the new model object. Returns ------- results Updated Results object, that includes results from both the original dataset and the new dataset. Notes ----- The `endog` and `exog` arguments to this method must be formatted in the same way (e.g. Pandas Series versus Numpy array) as were the `endog` and `exog` arrays passed to the original model. The `endog` (and, if applicable, `endog_quarterly`) arguments to this method should consist of new observations that occurred directly after the last element of `endog`. For any other kind of dataset, see the `apply` method. This method will apply filtering to all of the original data as well as to the new data. To apply filtering only to the new data (which can be much faster if the original dataset is large), see the `extend` method. See Also -------- extend apply rrS7Cannot append data of a different dimension to a model.r)refit fit_kwargscopy_initializationr4) rrr/rsr\r`rarjrr) r9rrrrrr4rrrars r:rzDynamicFactorMQResults.appendss@"1!@!@ ?"$%( $4$9%++a.q tzz33 34::---)* *%4 !w~ E: 3#9E>DE Er<c 2tj||\}}t|jdk(r|jdnd}||jj k7s||jj k7r td||d<t|$|f||d|S)a Recreate the results object for new data that extends original data. Creates a new result object applied to a new dataset that is assumed to follow directly from the end of the model's original data. The new results can then be used for analysis or forecasting. Parameters ---------- endog : array_like New observations from the modeled time-series process. endog_quarterly : array_like, optional New observations of quarterly variables. If provided, must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. fit_kwargs : dict, optional Keyword arguments to pass to `filter` or `smooth`. retain_standardization : bool, optional Whether or not to use the mean and standard deviations that were used to standardize the data in the current model in the new model. Default is True. **kwargs Keyword arguments may be used to modify model specification arguments when created the new model object. Returns ------- results Updated Results object, that includes results only for the new dataset. See Also -------- append apply Notes ----- The `endog` argument to this method should consist of new observations that occurred directly after the last element of the model's original `endog` array. For any other kind of dataset, see the `apply` method. This method will apply filtering only to the new data provided by the `endog` argument, which can be much faster than re-filtering the entire dataset. However, the returned results object will only have results for the new data. To retrieve results for both the new data and the original data, see the `append` method. rrSrr)rr4) rrr/rsr\r`rarjrrm) r9rrrr4rrrars r:rmzDynamicFactorMQResults.extendsf"1!@!@ ?"$%( $4$9%++a.q tzz33 34::---)* *%4 !w~ E(#9E=CE Er<c |jj|f|||d|} |r:tjj |j } | | j _|j| ||} | S)a# Apply the fitted parameters to new data unrelated to the original data. Creates a new result object using the current fitted parameters, applied to a completely new dataset that is assumed to be unrelated to the model's original data. The new results can then be used for analysis or forecasting. Parameters ---------- endog : array_like New observations from the modeled time-series process. k_endog_monthly : int, optional If specifying a monthly/quarterly mixed frequency model in which the provided `endog` dataset contains both the monthly and quarterly data, this variable should be used to indicate how many of the variables are monthly. endog_quarterly : array_like, optional New observations of quarterly variables. If provided, must be a Pandas Series or DataFrame with a DatetimeIndex or PeriodIndex at the quarterly frequency. refit : bool, optional Whether to re-fit the parameters, using the new dataset. Default is False (so parameters from the current results object are used to create the new results object). fit_kwargs : dict, optional Keyword arguments to pass to `fit` (if `refit=True`) or `filter` / `smooth`. copy_initialization : bool, optional Whether or not to copy the initialization from the current results set to the new model. Default is False. retain_standardization : bool, optional Whether or not to use the mean and standard deviations that were used to standardize the data in the current model in the new model. Default is True. **kwargs Keyword arguments may be used to modify model specification arguments when created the new model object. Returns ------- results Updated Results object, that includes results only for the new dataset. See Also -------- statsmodels.tsa.statespace.mlemodel.MLEResults.append statsmodels.tsa.statespace.mlemodel.MLEResults.apply Notes ----- The `endog` argument to this method should consist of new observations that are not necessarily related to the original model's `endog` dataset. For observations that continue that original dataset by follow directly after its last element, see the `append` and `extend` methods. )rrr4)rr)r\r6rr? from_resultsrr_apply) r9rrrrrrr4rr5rAr#s r:applyzDynamicFactorMQResults.applys}xdjju)o/>6L)"() !00==##%D%)CGG "kk#UzkB r<c  |j} |d}||jj}|jj|} d} g} t|dd}t|dd}|#|jdk(r| d|j gfgz } t %||||||xr|||| | |  }d }|stj|jjdd| jd d f| | j } |jd }d }| j"r$|j$| j&d|d<|d z }|j$| j&d|d<|j(| d}||j+t,||< |j.dd| dfjd|j.dd| df<d|j0_t5j6| j8}t;| j<D]=}|Dcgc]}|| jj>|vr|!}}d|j.||f<?|jA}|jB}|j(jE}d}d}tG|||tH|}|jJjM|||d z }d }d }t;tO| jjPD]R}| jjP|}||j8z }|jjR} g}!t;|jTD](}|!|jD"cgc] }"d|d zd|"c}"z }!*tj| |jV|jXd f|j|! }d|j0_ |jd}|jjZ}#|j8d k(r |#||f|d<nE|j|d<t;|j8D]}|#||||zf||j|< |j(|j8 d}||j+t,||< |j.dd|j8 dfjd}$|$|j.dd|j8 df<|jA}|jB}|j(jE}d}d!|}tG|||tH|}|jJjM|||d z }|}U|S#t$r|j!d}YMwxYw#t$r>|j.dd| dfj!d|j.dd| df<YwxYwcc}wcc}"w#t$r|j!d}YwxYw#t$r4|j.dd|j8 dfj!d }$YmwxYw)"a Summarize the Model. Parameters ---------- alpha : float, optional Significance level for the confidence intervals. Default is 0.05. start : int, optional Integer of the start observation. Default is 0. title : str, optional The title used for the summary table. model_name : str, optional The name of the model used. Default is to use model class name. Returns ------- summary : Summary instance This holds the summary table and text, which can be printed or converted to various output formats. See Also -------- statsmodels.iolib.summary.Summary NzDynamic Factor Results)rDrrrz EM Iterations) rrrQrJdisplay_paramsdisplay_diagnosticsredisplay_max_endogextra_top_leftextra_top_rightrSr,rrc d|zSNz%.2frOrs r:rXz0DynamicFactorMQResults.summary..s &1*r<c d|zSrrOrs r:rXz0DynamicFactorMQResults.summary..s vzr<rbz idiosyncratic: AR(1)rzvar.c |dSNz.2frOrs r:rXz0DynamicFactorMQResults.summary.. 3r<c |dSrrOrs r:rXz0DynamicFactorMQResults.summary..rr<zFactor loadings:rzObservation equation:rTrtrNc d|zSrrOrs r:rXz0DynamicFactorMQResults.summary..s fqjr<c d|zSrrOrs r:rXz0DynamicFactorMQResults.summary..s 6A:r<z error variancez error covariancec |dSrrOrs r:rXz0DynamicFactorMQResults.summary.. QsGr<c |dSrrOrs r:rXz0DynamicFactorMQResults.summary..rr<zTransition: Factor block ).r\rLrFrrrrrjr}rrrrr.rar_rdrbrr rr*r+rr\r[r>r@r0rkrarrvrlrr"r$rbrcr/rrr1r,r*r)&r9rrrQrJrrdisplay_params_as_listrerr5rrrrrrjrkrk_idio cols_to_cast base_ilocrrrrmrnrorprrrYrA lag_namesr[rformatted_valsrs& r:rjzDynamicFactorMQResults.summaryOsE8jj =,E  //Jjj11)2+ dM48 t^T:  # (;(;t(C k6F6F5G2I JK KO'/uEj*E/E 3!5/)? "L%<<##**1cff\.BA+EF!3+;+;=D ;xx 45 F$$KK': ;<./! ;;svv.A'BCDL<<1L!%l!3!:!:6!BD  *)-1vgh;)?)C)C(** !fWX+& 1DJJO #--0I3;;' )#,AaCFF$<$ %),%7%>%>v%F\"%)YYq5??2B2C/C%D%H%H,&N3A !eoo-../'')"kk $ 3 3 5 # 3A37# &e5%%h6A g3 jI" ;}}%9: ;"" *)-1vgh;)?)H)H(** !fWX+& *A8"B&?==)=>D?"&%)YYq5??2B2C/C%D%M%M,&NsU7T8:U $V$ V) 2V./0W8UUAV! V!.W  W 9X  X r)rN)rNNNN) NNF predictedFTNNNN) NNNNNNNFNFrNT)r NT)NFNTT)NNT)NNFNFT) g?NNNTFFNr)rJrKrLrMr;rNr)rr rrrrrrrmrrjrrs@r:r9r9 s?%%N8D04ZxKKB9EDH=AM^B&*@EDBF@E%)GREI9>CG"#ggr<r9)DrMstatsmodels.compat.pandasrr collectionsrwarningsrnumpyr>pandasr} scipy.linalgrrr statsmodels.tools.datar statsmodels.tools.validationr statsmodels.tools.decoratorsr #statsmodels.regression.linear_modelr +statsmodels.genmod.generalized_linear_modelrstatsmodels.multivariate.pcar"statsmodels.tsa.statespace.sarimaxr)statsmodels.tsa.statespace._quarterly_ar1r#statsmodels.tsa.vector_ar.var_modelrstatsmodels.tools.toolsrrstatsmodels.tsa.tsatoolsrstatsmodels.tsa.statespacerr statsmodels.tsa.statespace.toolsrrrrrr*statsmodels.tsa.statespace.kalman_smootherrrr statsmodels.base.datar!statsmodels.iolib.tabler"statsmodels.iolib.summaryr#!statsmodels.iolib.tableformattingr$rir&rQMLEModelr MLEResultsr9rOr<r:r"s =#;;3173;,6B3)4+?))@@,/-8^/$^/Bj Dj ZI)h''I)XRqX00qr<