BUILD — SYSTEMATIC TRADING RESEARCH
A system designed to prove its own ideas wrong
Paper accounts, daily bars, US equities and ETFs, three risk profiles in parallel. The stated goal for the first six months is explicitly not profit — it's a research pipeline that reliably identifies when a strategy is garbage. No returns are claimed here and none should be inferred. This is not financial advice; it's an engineering project about self-deception.
Personal research build. Paper accounts only. Not financial advice.
- 1
- evaluation of the final holdout, ever
- 3
- risk profiles validated in parallel
- 90-day
- paper run: fix bugs, never parameters
- Long only
- no shorting, no leverage, no margin
Why this is on a marketing portfolio
Because it's the same skill, with the excuses removed. Marketing analysis and strategy backtesting fail identically: you find a pattern in data you've already looked at, you convince yourself it's real, and you spend money on it. The difference is that a trading system tells you within months, in cash, and a marketing attribution story can flatter you for years. So I built the environment where the discipline is enforced by code rather than by intention, to learn what that actually requires.
The invariants
These are the rules the codebase enforces on itself, and they're all defences against a specific way of fooling yourself. Signals computed on one bar execute at the next bar's open — asserted in the engine, because the alternative is trading on information you didn't have yet. All time access goes through an injected clock and never a real-time call, because in a backtest a real-time call returns today's date and produces beautiful, meaningless results. Strategies emit target weights, never orders. Validation raises rather than warning and continuing, because a warning in a log is a warning nobody read. The risk layer can only reduce exposure, never increase it. Reconciliation mismatches halt rather than auto-correct. Every backtest appends to a trial registry automatically, with no manual step to forget — so the count of things I tried is honest, which is the only thing that makes a good result meaningful. And the one I like most: the final holdout is evaluated exactly once, and the guard that enforces it has no override flag. Not a flag I've promised not to use — no flag.
The finding that shapes it
A study of 888 real strategies with live out-of-sample results found that backtest Sharpe ratio explained under 2.5% of the variance in live Sharpe. Annual backtest return correlated negatively with live results. What did predict out-of-sample performance was volatility, maximum drawdown and portfolio-construction characteristics. So the ranking criterion is deliberately counterintuitive: strategies are ranked on volatility against target and on drawdown, not on the headline number everyone actually looks at. Building the thing that ignores the impressive number is the entire exercise.
What it proves
That I know the difference between a result and an artefact, and that I'll design a process that can tell me I'm wrong before I spend on it. That instinct is worth more in a media budget review than in a trading account.
Stack