Guides

Backtests are optimists. They replay one price history, in which nobody noticed your orders and the crash happened to somebody else. These guides test the parts a backtest leaves out, from a liquidity crisis and the cost of your own orders to a language model that may already know how 2020 ended, and a decision engine that never once said recession until we rewrote its options.

Four of them also come as demos, if you'd rather watch the hosted app do it.

• Stress testing a trading bot against a market crash

  Run one trading bot through a liquidity crisis, a rate shock and a recession on markets nobody has seen, and find out which crash breaks its drawdown limit.

• Testing an LLM trading agent without look-ahead bias

  Test an LLM trading agent on markets no model has seen, seal the seeds before the run, and record every model answer so anyone can replay the result.

• Skill or luck in a trading strategy

  One good backtest proves very little. Run two strategies on 40 generated markets, read the paired result, and see how a single market can crown either one.

• Measuring the market impact of your own orders

  Price every fill against the market where your order was never sent. See how cost grows with order size, and how much patience saves on a large order.

• Backtests and simulated markets

  A backtest answers one question well. Three it can't answer, about your own orders, luck and what-ifs, asked of a simulated market with the code to run.

• A risk gate for a trading bot in CI

  Put a risk gate in CI: twelve simulated markets, one drawdown rule, and a red build when a pull request makes the trading bot riskier than you agreed.