MatheLinuxQuant research · software · education
Product · Early access
An independent test of whether a strategy's reported performance reflects a real, repeatable edge, or is better explained by luck, overfitting, data errors, or unrealistic assumptions. Built for trading firms, investment firms, and asset managers that need evidence before committing capital.
We sell a process and its evidence. We do not sell trading signals, predictions, or investment advice.
How it works
A trade log or return series, and how many variants were tried before this one. Code is optional.
Tests and pass/fail criteria are fixed and hashed before anything runs. The engine runs them in isolation and a reviewer checks the result.
One of four verdicts, the findings behind it, and a reproducible report. The full evidence is there when you want it.
The problem
Research can now generate thousands of backtests quickly. When many variants are tried and the best is kept, its history is biased upward by selection alone. Other common failures:
Verdicts
Methodology
Tests and pass/fail criteria are fixed and hashed before any test runs, so nothing is chosen after the results are seen.
Timestamp alignment, gaps and bad ticks, survivorship, roll and corporate-action effects, fills that could not have happened, and returns that are too smooth to be real.
Walk-forward and purged cross-validation with embargo, so performance is measured on data the strategy was not fitted to.
Probabilistic and deflated Sharpe ratios, minimum track-record length, backtest-overfitting probability, and data-snooping tests such as White's Reality Check.
Doubled and tripled costs, execution delay, nearby parameters, sub-periods, and removing the best trades or days.
Whether returns are alpha or hidden beta, carry, or tail risk, and whether the regime the edge depends on can be recognised in real time.
Every run records its data hash, configuration hash, engine version, and random seeds, so any report can be regenerated exactly and checked independently.
Confidentiality
You provide results only. Your code is never seen; tests that need code are marked not applicable rather than skipped silently.
You also provide code and data, which allows look-ahead truncation tests and precise cost modelling, under stricter isolation.
Data is isolated per engagement and encrypted at rest. Jobs run in isolated containers without network access. Retention periods are set by contract, and an on-premises option is available.
Early access
Tell us who you are and what you trade. We will reply to arrange a scoping call.
