RAM-FinGAN

Regime-aware extension of Fin-GAN for stock--ETF excess-return forecasting.

This project keeps the original Fin-GAN stock--ETF forecasting setup and adds market-regime features built from ETFs, interest rates, volatility, and credit-spread proxies.

Reference paper: Fin-GAN: Forecasting and Classifying Financial Time Series via Generative Adversarial Networks

Forecasting Target

p(y_{t+1} | C_t, R_t)

where:

• C_t is the historical stock--ETF return window;
• R_t is the market-regime state;
• y_{t+1} is the next-period stock--ETF excess return.

1. What Is Added

Compared with the original Fin-GAN setup, this repository adds:

• market-regime features from broad ETFs, sector ETFs, Treasury yields, VIX, and credit-spread proxies;
• regime-factor models for compressed market-state conditioning;
• transaction-cost, smoothing, ticker-level, period-level, and regime-level evaluation;
• cleaned result figures for GitHub display.

Implemented model variants:

• baseline LSTM;
• RAM-LSTM with direct raw market features;
• regime-factor LSTM;
• RAM-FinGAN v1;
• RAM-FinGAN v2 with supervised pretraining;
• RAM-FinGAN v3 with economic conditioning;
• regime-economic LSTM v4.

The current strongest model is the regime-factor LSTM. The GAN variants are retained as ablation studies.

2. Data

Raw market data are not included in this repository.

CRSP/WRDS data should be downloaded from CRSP on WRDS.

Recommended WRDS path:

CRSP -> Annual Update -> Legacy Data - Stock / Security Files -> Daily Stock File

Sample period:

2000-01-01 to 2021-12-31

Expected local raw-data layout:

data_raw/
├── crsp/
│   ├── Stocks-data.csv
│   ├── ETFs-data.csv
│   └── Market-ETFs-data.csv
└── external/
    ├── VIXCLS.csv
    ├── DGS10.csv
    ├── DGS2.csv
    ├── DGS3MO.csv
    └── BAMLH0A0HYM2.csv

External series:

• VIXCLS
• DGS10
• DGS2
• DGS3MO
• BAMLH0A0HYM2

Note: historical access to BAMLH0A0HYM2 through FRED may be limited depending on the data interface. In that case, use the ICE source directly or replace it with a documented credit-spread proxy such as BAA10Y or an HYG-LQD spread proxy.

3. Main Results

Cleaned result figures are stored in:

results_figures/

3.1 Overall Model Comparison

The strongest zero-cost test performance is obtained by the regime-factor LSTM.

Model	Direction Accuracy	Mean PnL (bp)	Test Sharpe	RMSE	MAE
regime-factor LSTM	51.62%	2.88	0.551	0.01199	0.00697
RAM-FinGAN v2	51.18%	2.26	0.454	0.01208	0.00707
RAM-FinGAN v3	51.26%	1.80	0.374	0.01214	0.00715
baseline LSTM	51.19%	1.57	0.303	0.01199	0.00697
naive RAM-LSTM	51.01%	1.20	0.231	0.01205	0.00701

Key observation:

Raw market features are not robust when directly concatenated with LSTM inputs.
Compressed regime-factor representations provide stronger out-of-sample performance.

3.2 COVID-Shock Performance

The regime-factor LSTM is especially strong during the COVID-shock period.

Model	COVID-Shock Sharpe
regime-factor LSTM	0.797
RAM-FinGAN v2	0.580
baseline LSTM	0.172
naive RAM-LSTM	-0.126

This supports the main motivation of the project: market-regime information is most useful when the market environment changes sharply.

3.3 Transaction-Cost and Deployment Results

The raw regime-factor LSTM has strong predictive performance but high turnover. Transaction costs expose this weakness.

Cost (bp)	Raw Sharpe	Smoothed Sharpe	Turnover Reduction
0	0.551	0.390	66.3%
1	0.387	0.279	77.4%
2	0.222	0.231	94.8%
5	-0.274	0.149	97.5%
10	-1.106	0.108	99.8%

Key observation:

The regime-factor signal is useful, but practical deployment requires turnover-aware position smoothing or no-trade bands.

4. Core Figures

Only the most important figures are shown below.

Overall Model Ranking

Cumulative Test PnL

Transaction-Cost Robustness

Position Smoothing Under Costs

5. Usage

Install dependencies:

pip install -r requirements.txt

Run the pipeline after preparing the raw data:

python src/01_clean_and_check_raw_data.py
python src/02_fix_market_state_features.py
python src/05_make_lagged_market_features.py
python src/03_build_ram_panel.py
python src/04_check_ram_panel.py
python src/06_train_lstm_vs_ram_lstm.py
python src/07_train_regime_factor_lstm.py
python src/08_train_ram_fingan_v1.py
python src/09_train_ram_fingan_v2_pretrain.py
python src/10_train_ram_fingan_v3_econ.py
python src/11_train_regime_econ_lstm_v4.py
python src/12_analyze_all_models.py
python src/13_robustness_transaction_cost_bootstrap.py
python src/14_position_smoothing_cost_aware.py
python src/15_final_summary_tables.py
python src/16_make_publication_figures.py

Important note on ordering:

05_make_lagged_market_features.py must be run before 03_build_ram_panel.py.

This avoids look-ahead bias by ensuring that market-state features are lagged before they are merged into the stock-level forecasting panel.

6. Repository Contents

Public repository contents:

configs/
results_figures/
src/
README.md
requirements.txt
.gitignore

Excluded from the public repository:

data_raw/
data_clean/tickers/
data_clean/ram_panel/
outputs/

These excluded folders may contain licensed CRSP data, intermediate files, trained models, logs, or large outputs.

7. Interpretation

The main empirical conclusion is not that GAN variants dominate.

Instead, the evidence supports a more nuanced claim:

1. Directly adding raw market-state variables is not robust.
2. Compressing market information into regime factors improves out-of-sample forecasting.
3. The regime-factor model is particularly useful during market-shock periods such as COVID.
4. GAN-style adversarial training is unstable in this small-signal financial forecasting setting.
5. Transaction costs expose a high-turnover weakness.
6. Turnover-aware position smoothing improves deployability under realistic costs.

Suggested research direction:

Regime-factor representation and turnover-aware deployment
for robust stock--ETF excess-return forecasting under market regime shifts.

The GAN variants should be interpreted as ablation studies rather than as the final main model.

8. Disclaimer

This repository is for research use only. It is not financial advice.