This repository contains empirical results in paper to estimate a general non-linear asset pricing model with a deep neural network applied to all U.S. equity data combined with all relevant macroeconomic and firm-specific information.
We compare our GAN model, with a simple forecasting feedforward network model labeled as FFN, the linear special case of GAN labeled as LS and a regularized linear model labeled as EN.
- Sharpe Ratio
| Model | SR (Train) | SR (Valid) | SR (Test) |
|---|---|---|---|
| LS | 1.80 | 0.58 | 0.42 |
| EN | 1.37 | 1.15 | 0.50 |
| FFN | 0.45 | 0.42 | 0.44 |
| GAN | 2.68 | 1.43 | 0.75 |
- Explained Variation
| Model | EV (Train) | EV (Valid) | EV (Test) |
|---|---|---|---|
| LS | 0.09 | 0.03 | 0.03 |
| EN | 0.12 | 0.05 | 0.04 |
| FFN | 0.11 | 0.04 | 0.04 |
| GAN | 0.20 | 0.09 | 0.08 |
- Cross-Sectional R2
| Model | XS-R2 (Train) | XS-R2 (Valid) | XS-R2 (Test) |
|---|---|---|---|
| LS | 0.15 | 0.00 | 0.14 |
| EN | 0.17 | 0.02 | 0.19 |
| FFN | 0.14 | -0.00 | 0.15 |
| GAN | 0.12 | 0.01 | 0.23 |