Methodology
How we validate a strategy so the verdict means something.
Survivorship-free (point-in-time)
We reconstruct the S&P 500 at each historical date, including companies that later left or went bankrupt. Without this, the backtest inherits a phantom return.
Net of costs
Every rebalance subtracts transaction costs. Over short horizons, cost is usually the wall that knocks the signal down.
Delisting-aware exit
If a name stops trading, we use its last available price — we don't make it vanish from the calculation.
The verdict
“Looks like it goes up” isn't enough. We measure the net mean per rebalance, its p-value via date-clustered bootstrap, the PSR and the % of positive years. Only if it beats chance and cost robustly: PASS.
Honesty: a real example
Live news sentiment has no signal. Here is its verdict, unvarnished:
Verdict: SENIAL significativa: +6.5 bps/trade, p=0.046