Point-in-time (PIT) data reflects what was actually known at a specific historical moment. That matters more than it sounds. When you test a strategy against the past, you're pretending to be an investor on that day, with only that day's information. Use anything that arrived later — a revised earnings number, an updated index list — and the test quietly cheats.
Here's a simple example. A company reports Q1 earnings in April, then restates them in July. If your database shows the restated figures for April, your backtest is trading on numbers nobody had in April. That's look-ahead bias — incorporating information that wasn't available when the decision would have been made — and it makes results look better than they honestly were. PIT data keeps the original figures exactly as they were first published, warts and all.
Look-ahead bias is any use of information that didn't exist at the decision point you're simulating. Some forms are obvious. Others are sneaky.
Timing is a classic one. A company's fiscal quarter ends March 31, but it doesn't actually report until early May. If your backtest lets a strategy act on those numbers on April 1, it's trading on figures that were still five weeks from being public. A good PIT database stores the date each number became known, not just the period it covers.
Revisions are another. Economic data like GDP and unemployment gets revised for months after the first release. Company fundamentals get restated. If your test sees the final, cleaned-up value, it knows things no real investor could have known at the time. The fix is always the same question, asked relentlessly: on this date, what did the world actually know?
Survivorship bias occurs when analysis only includes companies that still exist today, ignoring those that went bankrupt, were acquired, or were delisted. This makes historical performance look better than it really was, because the failures have been quietly erased.
Picture a study of stock returns since 2000 that only includes currently active companies. It would exclude every firm that failed during the dot-com bust and the 2008 financial crisis — precisely the stocks that did the most damage to real portfolios. The survivors, by definition, are the ones that made it. Averaging only their returns is like judging a marathon by interviewing the finishers.
Rigorous analysis has to include delisted securities, all the way through their final trading days. That's less flattering to most strategies. It's also the truth.
Here's the most common way survivorship bias sneaks into backtests: testing a strategy on today's S&P 500 members going back twenty years.
Think about what that universe really is. Every company in it survived, and most thrived — that's how they earned their spot. Meanwhile the test excludes Enron, Lehman Brothers, and every other name that was in the index back then but blew up or faded away. The S&P 500 typically swaps out 20 to 30 companies a year, so over two decades the membership churns enormously.
It's a double cheat. You dodge the losers a real investor could have owned, and you 'discover' future giants years before the market crowned them — a 2004 test that includes Netflix only because Netflix later became huge is peeking at the answer key. This alone can inflate a backtest's returns by a few percentage points a year, sometimes more. The honest fix: use the index membership as it stood on each historical date.
Keeping dead companies in your data isn't enough — you also have to handle how they died.
When a stock delists, something specific happens to shareholders. In a bankruptcy, the equity often goes to zero or close to it. In an acquisition, holders receive cash or shares, sometimes at a nice premium. A dataset that simply stops tracking a stock at its last listed price can miss the final chapter, and the final chapter is often the whole story. A test that 'sells' a collapsing stock at its last quote before delisting may be assuming an exit that no real investor could have gotten.
Careful databases record delisting returns — the actual final outcome for shareholders — so a simulated portfolio takes the same hit, or gets the same buyout check, that a real one would have.
None of this means backtesting is broken. It means the data has to be built the way history actually happened: index members as of each date, delisted stocks included with their final returns, fundamentals stamped with the day they became public.
That's the standard Well Street builds on — strategies are tested against decades of real market history using point-in-time data, so the past you're testing is the past investors actually lived through, not a tidied-up version of it.
One honest caveat, always: even a perfectly clean backtest describes the past, not the future. A strategy that would have worked for thirty years can still disappoint for the next ten. Point-in-time data removes false confidence; it can't manufacture real certainty, because nothing can. What it gives you is a fair test. In a field full of flattering mirrors, a fair test is worth a lot.
Previous: ESG Investing · Next: Backtesting Strategies · Financial glossary
← Back to all investing concepts