Backtesting has a blind spot it can't see past
A backtest asks: if I had used this rule on past data, how would it have done? It sounds rigorous, and the arithmetic usually is. The problem is upstream of the arithmetic — the person building the test already knows how the story ends. They've seen which stocks crashed, which sectors ran, which months were strong. Even with real discipline, it's extraordinarily easy to shape a rule, consciously or not, around a pattern you can already see in the data you're testing it on.
This isn't a hypothetical risk. It's the default failure mode of almost all retail-facing "backtested" claims — the test and the discovery happen on the same data, in the same sitting, with the ending already known.
Point-in-time data closes that gap by construction
The fix isn't more discipline — discipline erodes under pressure. The fix is a data structure that makes the mistake impossible rather than merely discouraged. Every record here is stored exactly as it was known at the time: a corporate action is dated by its real announcement date, not the date it was typed into a database; a price is what was actually observed that day, never quietly revised because a later correction came in.
Concretely: if a company announced a bonus share issue on a Tuesday, a research test dated the prior Monday cannot see it — not because of a rule that says it shouldn't, but because the record simply doesn't exist yet at that point in the data. That's the difference between "we tried not to look ahead" and "there was nothing to look ahead at."
Then the harder part: grading it forward, not backward
Point-in-time data only protects the test. It doesn't prove the idea works. For that, a call has to be logged before its outcome is known, left alone, and graded automatically once its forward window closes — against a random baseline, not a comparison chosen after the fact. No re-running with hindsight. No quietly dropping the ones that didn't work out. The full five-step process is on the Methodology page.
This is a much higher bar than backtesting, and it's slower — a forward-graded record takes real calendar time to accumulate, there's no way to rush it. That's the tradeoff: a backtest can "prove" almost anything by next Tuesday; a forward record can only prove what it's actually lived through.
What this buys you, and what it doesn't
None of this guarantees a strategy works. Point-in-time data and forward grading don't create an edge — they just guarantee the test wasn't cheating. Most ideas tested this way don't survive it; see the Journal for what that record actually looks like, including the findings that didn't hold up.
This is a description of process, not a claim about outcomes. Nothing here is investment advice — see the FAQ.