Macroeconomic Forecasting Benchmark: Deep Learning and Econometrics

This repository contains the source code for a research project evaluating the efficacy of deep learning methods (Neural ODEs and RNN) against standard econometric benchmarks for macroeconomic forecasting.

The project addresses two distinct multivariate forecasting scenarios: the US data (FRED-MD) and a custom panel based on international data (OECD MEI).

Results

Project Structure

• main_fred.ipynb: Experiments on US data (FRED-MD) with large number of varaibles.
• main_mei.ipynb: Experiments on OECD data. Focuses on panel structures and cross-country heterogeneity.
• model/: Model implementations (Latent Neural ODEs, Neural CDEs, RNNs, DFM, VAR, Ridge Regression, Random Forest).
• mei/: Folder of the MEI dataset and its preprocessing.
• fred/: Folder of the FRED-MD dataset and its preprocessing.
• evaluator.py: The evaluation engine handling Walk-Forward Validation (backtesting) and statistical hypothesis testing (Diebold-Mariano, Friedman).
• dataloader.py: Data management pipeline ensuring strict temporal splitting to prevent look-ahead bias.

Installation

This project requires Python 3.9+ and relies heavily on the JAX ecosystem.

pip install -r requirements.txt

Usage Example

To run a comparative backtest on panel data:

from dataloader import MacroDataLoader
from evaluator import UniFreqPanelMacroEvaluator
from model.lagged_model import LaggedModel
from sklearn.linear_model import Ridge

# 1. Load and process data
loader = MacroDataLoader(df, config={'target_col': 'GDP', 'date_col': 'Date', 'features': [...]})

# 2. Define model builders (seed -> model)
models = {
    'Ridge': lambda seed: LaggedModel({'base_estimator': Ridge(), 'lags': 4}),
    'Benchmark': lambda seed: LaggedModel({'base_estimator': Ridge(), 'lags': 1})
}

# 3. Initialize Evaluator
evaluator = UniFreqPanelMacroEvaluator(loader, models, horizon=12, metric='rmse')

# 4. Run Walk-Forward Validation
results = evaluator.run_backtesting(end_train_date_list=backtest_dates)

# 5. Statistical Analysis
evaluator.print_significance_tests(results)

Note for latter project

• Create an orchestrator that only returns prediction series
• Create a class to manage significance tests, calculate errors, etc.
• Create a class for plots
• Request non-standardized variables from the model
• Let the model ask for the data ? Need extra horizon for DL methods (eg NODE) to stabilize learning when horizon is short ~1