Monolith Research
The Gate.
The test a strategy must survive to reach capital. It is deliberately hard to pass — hard enough that it retired both of the strategies their author most believed in.
Every candidate the firm builds is judged by the same open-source promotion gate before anything is deployed. The point of the gate is not to confirm good news; it is to make good news survive a battery of tests designed to expose it as luck, overfitting, or borrowed market beta. Most ideas do not survive — and the record of what did not is kept in public.
A backtest is the easiest thing in finance to make look good. Search enough configurations, flatter the costs, pick a kind window, and almost any idea will print an attractive Sharpe. The gate exists because the firm assumes its own results are wrong until a sequence of deliberately unkind tests says otherwise.
The sequence is enforced by design. No candidate moves from in-sample optimisation straight to capital. It is deflated for the size of the search, tested for overfitting, stressed under a pessimistic cost model, stripped of market beta, and finally checked once on a window it has never seen. Each gate requires a written record — and the written record is what makes a later result interpretable instead of merely impressive.
- 01Probabilistic Sharpe (PSR)
- The probability that the true Sharpe is above zero, given how long the record is and how non-normal its returns are. A short, fat-tailed record needs a much higher headline Sharpe to clear the same bar.
- 02Deflated Sharpe (DSR)
- The PSR, corrected for the number of variants searched before this one looked good. The more configurations tried, the higher a Sharpe must be to mean anything. This is the test that kills the most results — a high Sharpe found after thousands of tries is mostly luck.
- 03Probability of backtest overfitting (PBO)
- Estimated by combinatorial cross-validation: the history is split many ways, and we count how often the in-sample best configuration lands below the median out of sample. A high PBO means the backtest is fitting noise.
- 04Minimum track-record length
- The length a record would need before its Sharpe is statistically distinguishable from zero. If the record is shorter than that, the honest verdict is not "fail" but "not yet".
- 05Cost-robustness
- Every result is re-run at the realistic transaction cost and at a punitive multiple of it. An edge that survives at one times cost but dies at three is recorded as cost-fragile, not deployable.
- 06Beta strip
- Market beta is removed before the verdict, so a strategy that simply rode a rising market is not mistaken for skill. What is left is the part of the return the market cannot explain.
- 07Blind out-of-sample holdout
- A final, pre-registered window the strategy has never seen. It is checked once, under the parameters fixed during validation. The signal behaves there, or it does not advance. One shot.
A gate that nothing fails is not a gate. This one retired both of the strategies their author most believed in, including one that headlined a Sharpe of 1.72 on a short window and held at 0.108 over the full 17.6 years out of sample. Retracting that number was not the embarrassing part of the record; it was the load-bearing part.
Every strategy the gate has retired is documented in full, with the real numbers, in the public Strategy Graveyard. The kill-record is what makes “it passed” mean something.
The gate is not a claim; it is code. Probabilistic and Deflated Sharpe, PBO by combinatorial cross-validation, minimum track-record length and the cost-robustness battery are published and covered by known-answer tests — runnable on any return series, not just ours.
github.com/MonolithResearch/quant-validationCorrespondence welcome — for research, technology, employment, and collaboration conversations.
bilal@monolithresearch.uk