New in v0.2.2: ability to get uncertainty intervals for predictions and predictions on synthetic vintages.
New in v0.2.0: ability to get feature contributions to the model and perform automatic hyperparameter tuning and variable selection, no need to write this outside of the library anymore.
R wrapper for nowcast_lstm Python library. MATLAB and Julia wrappers also exist. Long short-term memory neural networks for economic nowcasting. More background in this paper in the Journal of Official Statistics.
Installing the library: Install devtools with install.packages("devtools"). Then, from R, run: devtools::install_github("dhopp1/nowcastLSTM"). If you get errors about packages being built on different versions of R, try running Sys.setenv(R_REMOTES_NO_ERRORS_FROM_WARNINGS="true"), then run the install command again. Note on updating the library: This R wrapper is not versioned. When there is a new version of library, update the Python library by running pip install nowcast-lstm==0.2.3 (substitute 0.2.3 with whatever the latest version is) from the command line, then from R rerun devtools::install_github("dhopp1/nowcastLSTM"). This should give you access to the latest functionality in R.
Installing Python: If you already have Python installed on your system, simply follow the install instructions from the nowcast_lstm Python library and point initialize_session to the path where your Python is installed later on.
If you don't have Python installed on your system, run the following commands in R once you've run devtools::install_github("dhopp1/nowcastLSTM"):
library(reticulate)
install_miniconda(path = miniconda_path(), update = TRUE, force = FALSE)
py_install(conda=miniconda_path(), "dill numpy pandas pmdarima torch nowcast-lstm", pip=TRUE)Example: nowcastLSTM_example.zip contains an R Markdown file with a dataset and more detailed example of usage in R.
Once all Python libraries are installed, run the initialize_session function in R each time you use the library.
library(nowcastLSTM)
# this function should be run at the beginning of every session. Python path can be left empty to use the system default
initialize_session(python_path = "path_to/python")
# if you installed Python via reticulate, use this. You may get a warning about requesting one path and getting another, but it should work regardless.
initialize_session(python_path = miniconda_path())
# use this to set Python location permanently
Sys.setenv(RETICULATE_PYTHON = "path_to/python")LSTM neural networks have been used for nowcasting before, combining the strengths of artificial neural networks with a temporal aspect. However their use in nowcasting economic indicators remains limited, no doubt in part due to the difficulty of obtaining results in existing deep learning frameworks. This library seeks to streamline the process of obtaining results in the hopes of expanding the domains to which LSTM can be applied.
While neural networks are flexible and this framework may be able to get sensible results on levels, the model architecture was developed to nowcast growth rates of economic indicators. As such training inputs should ideally be stationary and seasonally adjusted.
Further explanation of the background problem can be found in this UNCTAD research paper. Further explanation and results in this UNCTAD research paper.
Given data = a dataframe with a date column + monthly data + a quarterly target series to run the model on, usage is as follows:
library(nowcastLSTM)
initialize_session()
# this command will instantiate and train an LSTM network
# due to quirks with using Python from R, the python_model_name argument should be set to the same name used for the R object it is assigned to.
model <- LSTM(data, "target_col_name", n_timesteps=12, python_model_name = "model") # default parameters with 12 timestep history
#model <- LSTM(data, "target_col_name", n_timesteps=12, n_models=10, seeds=c(1:10), python_model_name = "model") # For reproducibility on a single machine/system, give a list of manual seeds as long as the n_models parameter. Reproducibility across machines is not guaranteed.
predict(model, data) # predictions on the training set
# predicting on a testset, which is the same dataframe as the training data + newer data
# this will give predictions for all dates, but only predictions after the training data ends should be considered for testing
predict(model, test_data)
# to gauge performance on artificial data vintages
ragged_preds(model, pub_lags, lag, test_data)
# save a trained model
# python_model_name should be the same value used when the model was initially trained
save_lstm(model, "trained_model.pkl", python_model_name = "model")
# load a previously trained model
# due to quirks with using Python from R, the python_model_name argument should be set to the same name used for the R object it is assigned to.
trained_model <- load_lstm("trained_model.pkl", python_model_name = "trained_model")To ease variable and hyperparameter selection, the library provides provisions for this process to be carried out automatically. See the example file or run ? on the functions for more information.
# case where given hyperparameters, want to select which variables go into the model
selected_variables <- variable_selection(data, "target_col_name", n_timesteps=12) # default parameters with 12 timestep history
# case where given variables, want to select hyperparameters
performance <- hyperparameter_tuning(data, "target_col_name", n_timesteps=12, n_hidden_grid=c(10,20))
# case where want to select both variables and hyperparameters for the model
performance <- select_model(data, "target_col_name", n_timesteps=12, n_hidden_grid=c(10,20))Produce estimates along with lower and upper bounds of an uncertainty interval. See the example file or run ? on the functions for more information.
interval_preds <- interval_predict(
model,
test_data,
interval = 0.95
)
ragged_interval_preds <- ragged_interval_predict(
model,
pub_lags,
lag = 2,
data = test_data,
interval = 0.95
)data:dataframeof the data to train the model on. Should contain a target column. Any non-numeric columns will be dropped. It should be in the most frequent period of the data. E.g. if I have three monthly variables, two quarterly variables, and a quarterly series, the rows of the dataframe should be months, with the quarterly values appearing every three months (whether Q1 = Jan 1 or Mar 1 depends on the series, but generally the quarterly value should come at the end of the quarter, i.e. Mar 1), with NAs or 0s in between. The same logic applies for yearly variables.target_variable: astring, the name of the target column in the dataframe.n_timesteps: anint, corresponding to the "memory" of the network, i.e. the target value depends on the x past values of the independent variables. For example, if the data is monthly,n_timesteps=12means that the estimated target value is based on the previous years' worth of data, 24 is the last two years', etc. This is a hyper parameter that can be evaluated.fill_na_func: a function used to replace missing values. Options arec("mean", "median", "ARMA").fill_ragged_edges_func: a function used to replace missing values at the end of series. Leave blank to use the same function asfill_na_func, pass"ARMA"to use ARMA estimation usingpmdarima.arima.auto_arima. Options arec("mean", "median", "ARMA").n_models:intof the number of networks to train and predict on. Because neural networks are inherently stochastic, it can be useful to train multiple networks with the same hyper parameters and take the average of their outputs as the model's prediction, to smooth output.train_episodes:intof the number of training episodes/epochs. A short discussion of the topic can be found here.batch_size:intof the number of observations per batch. Discussed heredecay:floatof the rate of decay of the learning rate. Also discussed here. Set to0for no decay.n_hidden:intof the number of hidden states in the LSTM network. Discussed here.n_layers:intof the number of LSTM layers to include in the network. Also discussed here.dropout:floatof the proportion of layers to drop in between LSTM layers. Discussed here.criterion:PyTorch loss function. Discussed here, list of available options in PyTorch here. Pass as a string, e.g. one ofc("torch.nn.L1Loss()", "torch.nn.MSELoss()"), etc.optimizer:PyTorch optimizer. Discussed here, list of available options in PyTorch here. Pass as a string, e.g."torch.optim.Adam".optimizer_parameters:named list. Parameters for a particular optimizer, including learning rate. Information here. For instance, to change learning rate (default 1e-2), passlist("lr"=1e-2), or weight_decay for L2 regularization, passlist("lr"=1e-2, "weight_decay"=0.001). Learning rate discussed here.
Assuming a model has been instantiated and trained with model = LSTM(...), the following functions are available, run help(function) on any of them to find out more about them and their parameters. Other information, like training loss, is available in the trained model object, accessed via $, e.g. model$train_loss:
predict: to generate predictions on new datasave_lstm: to save a trained model to diskload_lstm: to load a saved model from diskragged_preds(model, pub_lags, lag, new_data, start_date, end_date): adds artificial missing data then returns a dataframe with date, actuals, and predictions. This is especially useful as a testing mechanism, to generate datasets to see how a trained model would have performed at different synthetic vintages or periods of time in the past.pub_lagsshould be a list of ints (in the same order as the columns of the original data) of length n_features (i.e. excluding the target variable) dictating the normal publication lag of each of the variables.lagis an int of how many periods back we want to simulate being, interpretable as last period relative to target period. E.g. if we are nowcasting June,lag = -1will simulate being in May, where May data is published for variables with a publication lag of 0. It will fill with missings values that wouldn't have been available yet according to the publication lag of the variable + the lag parameter. It will fill missings with the same method specified in thefill_ragged_edges_funcparameter in model instantiation.gen_news(model, target_period, old_data, new_data): Generates news between one data release to another, adding an element of causal inference to the network. Works by holding out new data column by column, recording differences between this prediction and the prediction on full data, and registering this difference as the new data's contribution to the prediction. Contributions are then scaled to equal the actual observed difference in prediction in the aggregate between the old dataset and the new dataset.model$feature_contribution(): Generates a dataframe showing the relative feature importance of variables in the model using the permutation feature contribution method via RMSE on the train set.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent