Point in time data is a record of what was known at each historical moment, including later revisions, corrections, and publication timing.
Point in time data is a record of what was known at each historical moment, including later revisions, corrections, and publication timing. In backtesting, that matters because a strategy should only be evaluated on information that would actually have been available when each decision was made. Using a cleaned current dataset can silently replace the historical information set with a revised version assembled after the fact. That changes signals, trade timing, and measured risk adjusted outcomes.
The core mechanism is straightforward. A point in time dataset stores observations together with their effective timestamps and revision history. A backtest then queries the dataset as of each decision time, not as of today. A cleaned current dataset usually collapses this history into a single latest value for each observation. That latest value may include revisions that were published later than the date being simulated. When those revisions influence model inputs, the backtest uses information that was not yet known. This is a form of look ahead contamination because the simulation benefits from future corrections rather than the original state of knowledge.
This issue is closely related to market microstructure and venue fragmentation in trading data. Sonar Sciences notes that a stock can trade on multiple venues and that combining those venue level records into a single feed can hide important details about where and when liquidity appeared. Their example shows that quote and trade patterns differ across venues and that aggregate views can obscure the true sequence of events seen by a strategy at execution time. The same general lesson applies to point in time fundamentals, macro releases, and other revised datasets. When the historical state is compressed into a cleaner modern view, the backtest can become less faithful to the information actually observable at each moment.
A useful way to think about the difference is to separate observation time from availability time. Observation time is when the underlying economic or market event occurred. Availability time is when that value became known to the strategy. Revisions introduce additional availability times because a first release can be followed by later updates. A point in time database preserves these distinctions. A cleaned current database often does not. If a strategy reacts to a macro series, an earnings field, or a derived ranking based on revised values, then the signal path in the backtest can diverge from the signal path that would have existed live.
Those divergences affect standard performance statistics. Returns can change because entries and exits move when signals shift. Volatility can change because the timing and composition of positions change. The Sharpe ratio can change because both average returns and volatility change. Sonar Sciences defines the deflated Sharpe ratio as a way to adjust an observed Sharpe ratio for multiple testing and non normal returns, with the goal of estimating the probability that the measured Sharpe exceeds a benchmark after accounting for selection effects. In practical terms, if cleaned current data makes signals look more stable or more predictive than they were in real time, then the observed Sharpe can be inflated before any deflation step is applied. A more faithful point in time reconstruction can therefore produce a more credible deflated Sharpe ratio because the underlying backtest is less contaminated by revised information.
The same logic extends to overfitting diagnostics. Sonar Sciences describes a backtest overfitting audit framework that evaluates the robustness of a strategy by examining how sensitive performance is to specification choices and by testing whether in sample success is likely to survive out of sample scrutiny. A strategy built on cleaned current data can appear more robust than it really is because revisions smooth away the uncertainty and noise present in the original information set. Point in time data does not eliminate overfitting, but it removes one avoidable source of false confidence. That makes any later audit more meaningful because the input data better matches the information conditions under which the strategy would have operated.
The claim that point in time data yields more reliable backtest outcomes than cleaned current data is well supported at the conceptual level by these sources. They support three key ideas. First, preserving the exact historical information set is necessary for realistic simulation. Second, revised or aggregated data can mask the timing and structure of what was actually observable. Third, robustness measures such as the deflated Sharpe ratio and overfitting audits are only as credible as the historical data used to generate the backtest.
For that reason, the strongest supported conclusion is narrow and methodological. Point in time data is the appropriate foundation for backtests when historical values can be revised or when timing of information arrival matters. It improves reliability by aligning the simulation with the information that was actually available at each moment. Cleaned present day data can obscure revisions and timing, which can distort signals and make downstream performance and overfitting diagnostics look more convincing than they should.
Covered in depth in the Cross-venue market data & signals pillar hub.