Every signal provider eventually quotes an accuracy number. Ours is 78% directional accuracy. The figure is meaningless until you know exactly what it measures and how it was calculated — so here is the honest version.
What it measures
Directional accuracy is the share of predictions that get the direction of the next move right: when the model says up, does price go up? It deliberately ignores magnitude. A model can nail direction 78% of the time and still be useless if it is right on tiny moves and wrong on huge ones — which is why accuracy is necessary but never sufficient.
Why the base rate matters
Markets are not a fair coin. Over many periods, stocks drift up more often than down, so a model that always predicts "up" might score 53% just by going with the drift. Any accuracy claim has to be compared against that base rate, not against 50%. An edge is the gap above the base rate, after costs.
A 78% number on ten trades is luck. On thousands of out-of-sample predictions, it is a track record.
The traps in accuracy claims
- In-sample testing: scoring the model on the same data it learned from inflates accuracy and predicts nothing about the future.
- Cherry-picked windows: accuracy measured only during a trending bull market falls apart in chop.
- Survivorship: testing on today's index members ignores the names that got delisted.
- Tiny samples: a great-looking number over a handful of trades is noise dressed as signal.
How we hold ourselves to it
Sintinel evaluates signals out-of-sample and tracks performance over time rather than freezing a flattering number on a landing page. Accuracy is paired with payoff ratio and calibration so the metric reflects real-world results, and it is allowed to move — up or down — as the data comes in. That is the difference between a methodology and a marketing claim.