S-003Mixed
Shallow neural models vs technical rules, after costs
Do small neural models beat classic indicator rules, and does either beat simply holding the stock once trading costs are paid?
Result
Neural models beat indicator rules on 447 of 498 names, but the best learned model beat simply holding on only 50 after costs.
- Neural beats indicator rules
- 447 / 498
- Best model (Logistic regression) beats holding, net
- 50 / 498
- Indicator rules beat holding, net
- 4 / 498
What it means
Beating technical rules is a low bar. The bar that matters to an investor is beating holding the stock, net of costs, and almost nothing clears it. That is why Stock Score is built to judge quality before capital rather than to time entries.
Locked setup
- Universe
- US large caps (498 S&P 500 names)
- Window
- Same walk-forward windows for every model
- Method
- Logistic regression, MLP and LSTM direction models scored against technical-indicator rules on the same names and windows. Long-only returns are taken net of trading costs and compared with staying long the whole time.
Breakdown
| Model | Accuracy % | Beat holding, net | Beat holding, risk-adj. |
|---|---|---|---|
| Logistic regression | 52.8 | 50 | 139 |
| LSTM | 52.3 | 30 | 144 |
| MLP | 51.5 | 12 | 71 |
| Technical-indicator rules | 50.4 | 4 | 35 |
Replay it
python scripts/shallow_neural_vs_technical.pyOutputs: data/us500_shallow_vs_ta.csv, data/us500_costed_v2.csv