When Statistical Significance Wasn't Enough

The hypothesis

Built on published research connecting sudden price shocks and trading volume to short-term reversal.

Stage 1 — Statistical test

Under the project's frozen, pre-specified criteria, the signal cleared the bar at two of the tested horizons.

t = 2.2110-day horizon (HAC)

t = 2.4515-day horizon (HAC)

This was, at the time, the strongest statistical result the project had produced.

Stage 2 — Real trading costs

A cost model was built from scratch using actual Indian brokerage, exchange, and statutory charges — not an assumed flat number.

82–125%of gross return consumed by costs

Across every tested variant, trading costs consumed between 82% and 125% of the gross return before any other adjustment.

Stage 3 — Concentration check

The next question was whether the return was broad or lucky.

5 of 77–116rebalance periods

395–2,052%of total cumulative gain

Across the tested variants, just 5 rebalance periods — out of 77 to 116 total — accounted for between 395% and 2,052% of the entire cumulative gain. Removing those five periods alone flips every variant net-negative.

Verdict

Not deployed Statistically significant, but not sufficiently robust for deployment. The signal was real in the statistical sense — it wasn't noise — but it depended on a handful of extreme periods and didn't survive contact with real trading costs at the project's capital scale.

Why this is the flagship result, not a footnote

Most of what makes a backtest misleading isn't the initial statistical test — it's stopping there. This is the one result in the project that cleared the first bar, and the value was in continuing to test it anyway.