What Numbers Runner is
A set of tools that read the market, and a rule that whatever they find gets measured before it gets published.
The method
Every signal is tested against a base rate - what would have happened anyway - not against zero. A pattern that precedes a 5% gain in a market that rose 5% has found nothing. Tests use non-overlapping windows, because 21-day forward returns sampled daily share 20 days out of 21, and that inflates significance dramatically. The bar also rises with the number of things tried: test enough indicators and one looks brilliant by arithmetic alone.
What that produced
Mostly nothing - which is the honest outcome, and the reason to trust the rest. Several home-grown signals ran backwards and had to be inverted in the tools that used them: price above its moving averages preceded -0.46% over the next 21 days (t = -4.40) across 413,946 stock-days. Turn-of-month, Santa Claus and buying the dip all measured at roughly zero. Fibonacci price levels reversed 31% of the time against 33% for deliberately arbitrary levels.
Two things survived: the variance risk premium, and the finding that essentially all of the equity premium arrives overnight rather than during the session. Neither was invented here. Both are published effects with a structural reason to exist.
Why the failures are published
Because a record that contains only winners is not a record. Signals are logged before the outcome is known, with a horizon fixed in advance, and scored when that horizon passes - whatever the answer. The record is here, losers included.
What this is not
Not advice, not a recommendation, and not a signal service. The tools describe what they measure and say where the evidence stops. Every decision is yours.
← Home