Backtesting bias

Learn how survivorship bias, look-ahead bias and overfitting distort stock backtests, what point-in-time data helps with, and what Tessera cannot guarantee.

Updated · 3 min read · All guides

Simulation conventions

Fills
Historical opening price, 10 bp per side
Dividends
Credited as cash on the ex-date
Daily disaster stop
Not applied in backtests
Inputs
Dated factor-panel slice per rebalance
Personal taxes
Not modelled

A historical simulation, not an account or a forecast.

A stock backtest applies rules to historical data. Its result is only as credible as the universe, the information available at each decision and the process used to choose the rules.

Three recurring problems are survivorship bias, look-ahead bias and overfitting. Point-in-time data addresses part of the problem, but it does not automatically make a historical simulation unbiased.

Survivorship bias: which companies are missing?

Survivorship bias appears when a backtest starts with companies that exist today and treats them as the entire historical opportunity set. Businesses that failed, merged or delisted can disappear from the test, changing both the choices and the losses represented.

Imagine testing a rule on today’s successful retailers without including retailers that closed during the period. Even accurate prices for the survivors cannot reconstruct the missing choices.

Historical membership, symbol changes, delisted prices and corporate-event treatment all matter. Finding a delisted ticker somewhere in a database is not proof that every historical failure is represented correctly.

Tessera’s approach: scoring reads dated factor-panel slices rather than simply applying today’s ratios to earlier prices. This preserves the available cross-section for each stored date. It does not establish complete delisted-stock coverage or correct cash proceeds for every acquisition and failure. Coverage needs to be assessed for the specific dates and names in a run.

Look-ahead bias: did the information exist yet?

Look-ahead bias uses information before it was available. Common examples include trading on an earnings report before release, applying a later restatement to an earlier decision, or generating a signal from a close and assuming an earlier fill.

A fiscal period-end is not the date investors learned the result. The release or filing timestamp is the important boundary. Prices, analyst estimates, ownership filings and sector classifications also need their own availability treatment.

Tessera’s approach: the engine reads the latest panel slice on or before the decision date. Factor materialization uses availability gates for supported inputs, and market context is read as of the relevant date. These controls reduce specific leaks; they do not prove every source contains a complete original-publication and revision history.

Some company classification inputs use current metadata. Missing prices and corporate events can also affect the simulation. We therefore do not claim historical classifications or a universal absence of look-ahead bias.

Overfitting: how were the rules selected?

Overfitting appears when a strategy is chosen because it fits the tested history unusually well. Trying many weights, filters and date ranges, then showing only the best combination, hides how much searching occurred.

A strategy can use correctly dated data and still be overfit. Point-in-time accuracy concerns information timing; it does not make a chosen model predictive.

Useful checks include reserving later data, evaluating multiple periods, testing neighboring parameter choices and recording how many alternatives were tried. Once a holdout period influences the next change, it is no longer an untouched holdout for that change.

Tessera’s limit: using the same scoring and construction engine across modes makes rule differences easier to inspect. It does not automatically turn every user backtest into an out-of-sample experiment. Preset selection and parameter changes remain research decisions whose history matters.

Questions to ask before trusting a backtest

  • Which dated universe and factor version did the run use?
  • Are failed and acquired companies represented throughout the tested period?
  • When did each financial input become publicly available?
  • How are absent prices, dividends and corporate events handled?
  • What execution prices and costs are assumed?
  • How many configurations were tested before this one was selected?
  • Which evaluation period remained untouched by that selection?
  • Which risk rules differ between historical and paper operation?

No single favorable answer settles the rest. Read Tessera’s backtest assumptions and limitations before using a simulation to compare configurations. The track record is a separate reporting surface and also needs its own dates and methodology understood. Compare plans for current research access.

Keep reading

Try it on a company

Open a company