On this page
The short answer
Can machine learning pick better stocks than Birch's plain rules? No. XGBoost predicted which of Birch's entries would win barely better than a coin flip. None of the 244 versions we tried passed our luck test.
Machine learning is the obvious next step for a stock-picking model. Instead of fixed rules, a model learns from ten years of NSE data which of Birch's entries turn into winners, and which stocks to prefer. XGBoost, a widely used tree-based learner, is the standard tool for this kind of data. We also tried simpler statistical models.
How did we test it?
- Three uses of the models:
- scoring Birch's own entries (keep only the ones the model likes, or put more into them);
- ranking all stocks;
- ranking stocks inside the strongest sectors.
- Three kinds of model:
- XGBoost;
- ordinary and ridge regression;
- Fama–MacBeth factor regressions, refitted every year.
- The inputs: price trend and momentum, volatility (how much a price jumps around), other price-position measures, delivery percentage, sector breadth, and the reaction to quarterly results.
- Honest training: each model learnt only from years before the ones it was tested on (walk-forward from 2019).
- 244 variants in all, each traded with Birch's market checks kept and also removed.
- The data: prices adjusted for splits, bonuses and demergers, all 2,008 stocks, to 28 October 2026. The models were retrained and every variant re-tested on this data on 2026-10-29.
- The checks:
- 6 start dates;
- the 5 best stocks removed;
- 2023 onwards on its own;
- the biggest 50 to 200 stocks removed;
- a luck adjustment across all 244 tries (the deflated Sharpe ratio, a score of return for the risk taken that allows for how many versions we tried), and the chance that the test on past prices was overfitted.
What did we find?
Rejected.
- Prediction was barely better than a coin flip. On Birch's entries, XGBoost's accuracy score (AUC) was 0.539, where 0.5 is a coin flip. A version forced to respect the obvious directions scored 0.553. Ordinary regression scored 0.526. The model overfitted, meaning it learnt noise from the past: in most test years its own stopping rule kept only 10 to 50 trees.
- Simple beat complex. For ranking stocks six months ahead, a plain linear factor model scored a rank correlation of 0.096, and 12-month momentum alone 0.088. XGBoost scored 0.054.
- The best trading variant did not survive the luck test. It combined a volatility filter with an XGBoost position-size tilt. Figures are averages over 6 start dates.
| Check | Birch | Volatility filter + XGBoost tilt |
|---|---|---|
| Without the 5 best winners | 25.4% | 26.5% |
| 2023 onwards, without the 5 best winners | 15.2% | 19.1% |
| Worst fall | 29% | 30% |
- After the luck adjustment: allowing for 244 tries, the deflated Sharpe of any variant's extra return over Birch was at most 0.04. We needed 0.95. The chance of an overfit result was 0.66.
What failed?
- The edge came from the volatility filter, not XGBoost. The filter alone did better than the filter with the tilt (27.5% against 26.5% without the 5 best winners). The filter itself turned out to be an artefact of missing data (see "Should a model hold only volatile stocks?").
- Without the biggest stocks (the top 100 to 200 removed), the XGBoost tilt lost its edge over plain Birch.
- A later learning-to-rank test (2026-10-28) gave the same answer. XGBoost's gain over momentum was −0.012 in 2019–21 and +0.025 in 2022–26. That is not statistically significant, so it could easily be luck (t 1.26).
- Removing the market checks cost most variants 2–10 points with the 5 best stocks removed, and deepened the worst fall to 30–43%.
When the outcomes are this noisy, a bigger model adds little skill. Simple trend measures carry most of what the learners found.
What are the limits?
- The price data holds only companies still listed today. That flatters every result here.