Blog › Method
Point-in-time data: why so many backtests cheat without knowing it
A strategy that beats the market year after year on paper and then fails in real life often has the same culprit: the backtest used information that did not exist at the time.
16 November 2026 · 5 minute read
Look-ahead bias
It means using, in a simulation, a figure that was not yet known on the simulated date. The obvious case is using a quarter's results on the day the quarter ends, when the company publishes them weeks later. That one is easy to avoid. The dangerous ones are those you cannot see.
The three invisible traps
1. Restated accounts
Companies correct their figures: they restate a year, change an accounting standard or reclassify a business they have sold. Most databases overwrite the old figure with the new one. That is right for someone looking today, but a backtest reading that database believes the figure corrected in 2024 was already known in 2022.
2. Companies that no longer exist
If the database only holds the companies listed today, it is missing the ones that went bust, were taken over or delisted. A backtest over that universe only picks among survivors, and survivors, by definition, did well. That is survivorship bias, and it almost always inflates results.
3. The date that is not the date
A figure has two dates: the period it refers to and the day it became known. Many databases only keep the first. Without the second there is no way to know what an investor had in front of them on a given day.
How we store our data
At Market Pi-Rates we never update a figure. Every morning we capture every company in five markets and store it as a dated snapshot that is never touched again. If a company corrects a figure, the correction goes into that day's capture and the previous one stays as it was.
- Since 20 August 2026 there have been 238 captures, and every working day adds a new one per market.
- Each capture is sealed with a fingerprint once it stops being the latest. If anyone changed a single past figure, the fingerprint would no longer match, and we check it before the data is published.
- Companies that delist are not deleted: they stay in the captures they were part of.
So the question "what was known about this company on 3 October?" has an exact answer, and it will be the same answer a year from now.
The honest limit
A history like this cannot be rebuilt backwards: either you stored it at the time or you do not have it. Ours starts in August 2026, so today it is good for following how companies change capture by capture, not for a ten-year backtest. It gets longer every day, and that is precisely its value.
How to query it
With Market Pi-Rates connected to your AI assistant, just ask:
"Show me how ASML's operating margin has changed in every capture since September."
"Which Spanish companies have seen their earnings per share turn around in the last month?"
The full history and the computed ratios come with the Navigator plan. Today's figures for all 25,000 companies are free.
See the plans