Quantitative research programme

Our own programme — machinery built to tell us when there is no edge

Sector: Quantitative research · Year: 2025–2026

Summary

An in-house systematic research programme. What we describe here is the part we are willing to describe: the tests. Every strategy must clear a single shared validation gate before it is allowed paper trading, and clear it again before it is allowed capital. How candidates are produced is not published, and this page does not describe it.

Challenge

The hard problem in systematic trading is not producing a high backtest. It is distinguishing a real edge from the best of N noisy draws. Each of our strategies previously carried its own private backtest with its own fill logic, its own fee constant and its own success criteria — which is exactly how a number reaches a capital decision without ever having been re-examined by anything that could disagree with it.

Approach

  • Made validation independent of research. The gate scores and does not tune, because anything that both searches and validates will eventually validate its own search.
  • Made the gate mandatory and unforgiving: it refuses to score a strategy that cannot reproduce its own reported result, and every strategy is judged on shared simulation, costing and statistics rather than on its own.
  • Deflated for the size of the search — clustered t-statistics, permutation tests, and deflated Sharpe ratios against a floor agreed before testing begins, with variations of one idea counted as one idea.
  • Modelled costs from real order-book depth at the size being traded and ran friction sensitivity on every candidate, because an assumed flat spread is the most common way a paper edge disappears.
  • Calibrated the gate itself against known signal and known noise, proving it admits the first and rejects the second, rather than trusting it because it was carefully written.

Outcome

A shared harness scoring against a ten-asset, 533,213-bar perpetuals history, and an archive of documented nulls that stops the same idea being rediscovered six months later. The programme runs on research and paper trading. No client capital is involved, and no strategy, signal or system design is published.

Stack

Python, NumPy, pandas, SciPy, Parquet, TypeScript

More work