The Signal · Issue 01 · Momentum Strategy
What happens when ML
actually survives the audit?
actually survives the audit?
Most AI momentum research fails one test: remove the lookahead bias and see what's left. We did. Six models. Walk-forward validated. Here's what survived — free for practitioners.
Get the Free Research One-Pagers
Two PDFs — results & methodology. No credit card. No account creation.
No spam. Unsubscribe anytime. New issues published monthly.
15.4%
CAGR
Gradient Boosting
Gradient Boosting
0.717
Sharpe Ratio
vs 0.385 SPY
vs 0.385 SPY
−22.6%
Max Drawdown
vs −23.9% SPY
vs −23.9% SPY
6
ML Models
All Tested
All Tested
45K+
Finance
Practitioners
Practitioners
Current Issue · Momentum Strategy · Walk-Forward Results
The result that survived the audit
Six ML models tested against SPY across 30 U.S. equities, weekly rebalancing, with walk-forward validation and embargo. The improvement is modest, stable, and economically believable — exactly what rigorous research produces. Gradient Boosting led the field. All models beat SPY on a risk-adjusted basis.
| Strategy | CAGR | Sharpe | Volatility | Max Drawdown | Final Value |
|---|---|---|---|---|---|
| SPY (Benchmark) | 9.5% | 0.385 | 16.9% | −23.9% | 131.7 |
| Gradient Boosting — LC | 15.4% | 0.717 | 17.4% | −22.6% | 153.7 |
| 30 liquid U.S. equities · Long/Cash · Weekly rebalancing · Walk-forward with embargo · No lookahead bias · Source: DHI Research | |||||
One-Pager 1 · Results
Strategy Performance
- Equity curves — all 6 models vs SPY benchmark
- Drawdown comparison — full underwater chart
- CAGR, Sharpe, volatility, max drawdown table
- Why the result is modest — and why that matters
- What most AI finance demos get wrong
One-Pager 2 · Methodology
Research Framework
- GPT responsibilities vs. human responsibilities
- 5 governance principles — the non-negotiables
- How leakage was detected and eliminated
- Traditional vs. GPT-augmented research — what changed
- Why GPT upgrades the researcher, not replaces them
"GPT collapses research friction — code, diagnostics, iteration that used to take weeks now takes hours. But humans retain control of credibility — data integrity, leakage detection, economic plausibility. The bottleneck in quant research has shifted from implementation to judgment."
Get the research. Free.
Two focused PDFs. No account. No password. Just the work.
Subscribe to The Signal →