Methodology · our own strategy, audited
Does our own trading bot actually work? Here’s the honest backtest.
We spend these posts poking holes in other people’s edges. Fair is fair — here’s ours, warts shown. The signal is real enough to trade on paper, and thin enough that we won’t pretend it’s more.
What the bot buys
Insider clusters, not politicians
The bot ignores Congress entirely — as we’ve shown, those disclosures are 45 days stale by law. It trades corporate insider clusters instead: two or more insiders at the same company buying their own stock on the open market, weighted so a CEO or CFO counts double, with a combined value over $250k. Form 4 is due within two business days, so the signal is days old, not weeks — fresh enough to act on. Trend and volatility filters (price above its 50-day average, VIX below 30) gate the rest.
The test
Split the history, don’t peek
We swept the thresholds across 2021–2025, then split them: fit intuition on the in-sample years (2021–23) and judge on the out-of-sample years (2024–25) the rules never saw. A signal that only works in-sample is a curve-fit. Here’s per-trade alpha vs the S&P at the $250k value floor, by how strict the cluster score is:
| Cluster score | In-sample α | Out-of-sample α | Verdict |
|---|---|---|---|
| ≥ 3 (loose) | −0.9% | +3.2% | fails in-sample |
| ≥ 4 | −0.1% | +4.1% | borderline |
| ≥ 5 (what we run) | +0.8% | +3.5% | positive in both |
Only the strict score ≥ 5 slice is positive in both periods. That’s the whole reason it’s our live threshold — not because it’s the highest number, but because it’s the one that didn’t fall apart out of sample.
Why we don’t oversell it
Small, and within noise
Two numbers keep us honest. The in-sample alpha at score ≥ 5 is only +0.8% — barely above zero. And the win rate is 44%: fewer than half the trades beat the market, so the positive average leans on a handful of winners, exactly the pattern we flagged in Congress. On this much data, a +3.5% out-of-sample edge is roughly one standard error from nothing.
An edge you can’t distinguish from luck in a backtest is a hypothesis, not a strategy. So we run it live, on paper, in public — and let the forward record settle it.
That’s why the bot paper-trades a simulated $100k book rather than real money, why an AI verdict and a routine-vs-opportunistic insider factor ride along as shadow signals that never gate a trade, and why every position and its outcome is logged. The backtest earns the strategy a live trial. Only the live trial can earn it your trust.
The fine print
- What alpha means here. Per-trade return minus the S&P over the holding period, 2021–2025, before costs and slippage.
- Survivorship + regime. Five years is one market regime (a long bull with two scares). An edge that lived here may not survive a different one — the out-of-sample split guards against curve-fitting, not against the future looking nothing like the past.
- Not advice. A paper strategy we run in the open. Reproducible from SEC EDGAR Form 4 data; the sweep is in the repo.
