This repository contains the code used in the Bank of England Staff Working Paper 984 An interpretable machine learning workflow with an application to economic forecasting by Marcus Buckmann and Andreas Joseph.

In the paper, we propose a generic workflow for the use of machine learning models to inform decision making and to communicate modelling results with stakeholders. It involves three steps: (1) a comparative model evaluation, (2) a feature importance analysis and (3) statistical inference based on Shapley value decompositions. The paper demonstrates each step by forecasting changes in US unemployment one year ahead.

The data set we use is FRED-MD, a US database of monthly macroeconomic measurements. We use the snapshot of August 2021. Please note that the first row of the data indicates the transformations of variables as suggested by the authors of the database (see McCracken & Ng, 2015).

The code published here is not intended as a stand-alone package. It can be used to reproduce the results of the paper. But as the proposed workflow is quite general, parts of it may be transferred to other applications. No warranty is given. Please consult the licence file.

Should you have any questions or spot a bug in the code, please send an email to [email protected] or raise an issue within the repository.

The file requirements/requirements.txt specifies the Python packages we used. We ran the experiments on a Windows 10 computer and set up the Python environment with the following commands:

conda create -n workflow python=3.9.7
activate workflow
pip install -r requirements.txt

We use Python to train and test the machine learning models and compute the measures of variable importance and use R to produce the figures and tables presented in the paper. The helpers folder contains helper functions and configurations:

• utils.py contains general helper functions, e.g. reading and writing data, or normalising and transforming variables
• utils_predict contains functions used to train and test the prediction models, e.g. it specifies the hyperparameter grids of the machine learning models
• utils_importance contains the functions to estimate variable importance.
• utils.R helper functions used when analysing the results in R.
• config.py contains some configuration parameters that are constant across all experiments.
• import_packages imports all Python packages used.

The script experiment.py is the main script. It reads the data and calls the run_experiments function, which loops through all the experiments. The parameters of the experiments are read from a yaml file (in folder setup_files) that is created with the script setup_epxeriments.py. In that script, the user can specify which prediction methods to test and can set parameters of the experimental set-up, such as the degree of winsorising, the type of cross-validation used for hyperparameter optimisation, or the training sample size. The user can either set these parameters to a single value (e.g. method = "Forest") or to a list of several values. In the latter case, all parameter combinations will be computed and the run_experiments will produce results for each of the experiments.

The run_experiments function computes the predictions of the model as well as the variable importance measures (Shapely values, permutation importance). For each experiment, output files with the predictions and variable importance measures are saved as .pickle files on the hard drive in the results folder. The name of these files are hash keys based on the parameter values of the experimental setup (such as method = "Forest").

After running the individual experiments, their results can be aggregated using the collect... scripts.

The script collect_performance.py appends the forecasting results of the individual experiments and saves them to a .csv file in the folder results/aggregated. Similarly, the scripts collect_importance_forecast.py aggregated the variable importance results of the forecasting experiments into a .csv. In the paper, we also estimate Shapley values using the out-of-bag approach. This means, we train the models on a bootstrapped sample of the data and estimate the Shapley values on those observatios not in the bootstrapped sample (i.e. the out-of-bag observations). The script collect_importance_out_of_bag.py collects these results.

The script shapley_regression.py reads the file containing the Shapley values of the forecasting experiments (that was compiled by collect_importance_forecast.py) and conducts statistical inference on the Shapely values using the Shapley regression method (see Joseph, 2019).

The script error_analysis.R reads the aggregated prediction results, computes the prediction errors, and plots time series of the observed and predicted response.

The script shapley_analysis reads the aggregated Shapley values and permutation importance values and produces the main variable importance figures shown in the paper.

This package is an outcome of a research project. All errors are those of the authors. All views expressed are personal views, not those of any employer.