How a market idea earns the right to be taken seriously

The division of labour: code counts, people decide

Every number in our research is produced by deterministic code that can be run twice and give the same answer: returns, rankings, trading costs, slippage, turnover, volatility, drawdown, Sharpe, Sortino, Calmar, exposure, statistical tests and calibration. Language models are used for what they are actually good at — reading source material, proposing a falsifiable hypothesis together with the null it must beat, writing a specification, and arguing against a conclusion. A model never calculates a result and never authorises anything. Promotion of any strategy is a human decision, recorded as a proposal with a name attached.

The sequence every hypothesis walks

Each stage is stored with its inputs, outputs, model, timestamps, cost and status. A stage cannot be skipped: the server refuses a transition whose predecessor has not passed, so a promising result cannot jump the queue to the front.

1. 1. Observation

   Write down what was noticed in the data or the world, before any explanation is offered.

   Written or reviewed by a person or agent

2. 2. Data quality check

   Check the data itself: gaps, duplicates, revisions, and whether there is enough history to say anything.

   Computed by code

3. 3. Hypothesis

   State a falsifiable claim together with the null hypothesis it has to beat, so failure is possible.

   Written or reviewed by a person or agent

4. 4. Strategy specification

   Turn the claim into a machine-readable specification: universe, features, execution model, costs, limits.

   Written or reviewed by a person or agent

5. 5. Code and unit checks

   Implement it in deterministic code and unit-test the arithmetic against known values.

   Written or reviewed by a person or agent

6. 6. In-sample backtest

   Run it over the testing history with commission and spread charged on every rebalance.

   Computed by code

7. 7. Walk-forward

   Fit on one window, trade the next, repeatedly, so no parameter benefits from knowing the future.

   Computed by code

8. 8. Locked holdout

   Read the reserved period once. It was locked before testing started and is never used for tuning.

   Computed by code

9. 9. Robustness

   Perturb parameters, cut the data into subsamples and regimes, and multiply the trading costs.

   Computed by code

10. 10. Red team

    Argue against the result on purpose: look-ahead, leakage, survivorship, too many trials, capacity.

    Written or reviewed by a person or agent

11. 11. Risk review

    Check exposure, concentration and loss limits against the stated risk policy.

    Written or reviewed by a person or agent

12. 12. Research report

    Write up what was found, what was not, and what remains unknown — including the failures.

    Written or reviewed by a person or agent

13. 13. Promotion proposal

    Ask a named person to decide. Nothing is promoted automatically, ever.

    Written or reviewed by a person or agent

The four tests that kill most ideas

Walk-forward, not hindsight

Parameters are fitted on one window and traded on the next, repeatedly, moving forward through history. A rule that only works when you already know how the period ended fails here.

A holdout that is genuinely locked

A period of history is reserved when a strategy version is created, before any testing begins. It cannot be changed afterwards through the admin screens or the API, and it is read once. If a conclusion falls apart there, it falls apart.

Costs charged at a realistic level

Commission and half-spread are charged on every rebalance, and the whole test is re-run with costs multiplied to see how much of the result was ever real. Findings that only exist at zero cost are recorded as failures, and kept as failures.

Counting how many times we tried

The number of variants tested is stored alongside the result, and used to compute a deflated Sharpe ratio and an estimate of the probability of backtest overfitting. Trying two hundred ideas and reporting the best one is not research; it is selection.

Adversarial review, on purpose

Before a research report is written, a red-team pass looks specifically for the ways we might have fooled ourselves: information used before it was knowable, labels leaking into features, a universe that quietly excludes what failed, too many trials, parameters that only work at one setting, fills that would not have happened, position sizes larger than the market could absorb, and results that depend on a single market regime. Any one of these can block a strategy from progressing, and the block is recorded rather than argued away.

What we openly do not have

• No paid market-data vendor, order-book feed or option-chain data. Our history is what we have recorded ourselves, and it is still short.
• No live brokerage connection. Nothing in this platform can place a real order, and that is a build-level fact, not a setting.
• No claim that a tested strategy will work in future. Testing narrows the range of plausible explanations; it never establishes one.

Evidence and sources

1. 1. Bailey & López de Prado, The Deflated Sharpe Ratio (2014) — Why a Sharpe ratio must be discounted for the number of variants tried before it means anything.
2. 2. Bailey, Borwein, López de Prado & Zhu, Pseudo-mathematics and financial charlatanism (2014) — The probability of backtest overfitting, and the CSCV procedure we use to estimate it.
3. 3. Harvey, Liu & Zhu, …and the Cross-Section of Expected Returns (2016) — The multiple-testing problem in published factor research; the reason we count trials openly.
4. 4. Novy-Marx & Velikov, A Taxonomy of Anomalies and Their Trading Costs (2016) — How much of a documented edge disappears once realistic trading costs are charged.