Evidence explorer: 632 designed backtests, refusals included, open to everyone

How do simple trading rules really do? The new public evidence explorer answers with a designed set of backtests rather than a selection of whatever users happened to run. Every rule is tested on fixed markets and fixed windows at the same stated costs, and the runs that were refused stay on the page with their reasons. You don't need an account to use it.

What's in it. 22 rules: the 17 distinct rules behind our 27 community templates (moving-average, RSI and breakout timing rules, and momentum, rotation, low-volatility, risk-parity and HRP rankings), plus 5 rules that combine conditions. They run on 11 fixed universes in the US, UK and Europe, each in one currency: single funds such as SPY, QQQ and the FTSE 100 ETF, and baskets of US mega caps, US sector ETFs, 10 FTSE 100 shares and 10 Euro Stoxx 50 shares. Each runs over 5, 10, 15 and 20 years ending 25 September 2026, and each template also runs on its own symbols. That makes 632 runs. Rules are only paired with universes they make sense on; a timing rule for one instrument isn't run on a basket, for example.

Same costs, same engine, never mixed. Every run uses the engine request the Strategy Agent would build for that prompt, with 1.0 basis point of slippage and 0.5 of commission per side, total return with dividends after withholding tax, and stamp duty on UK share purchases. Each run is compared with an equal-weight buy-and-hold of its own symbols, bought once. Each generation of the corpus is labelled with its engine version and cost model, and runs from different generations are never shown on one chart.

Refusals stay in. The number of runs shown against the total is always on screen. In the generation published on 28 September, 602 of the 632 ran. The other 30 are listed with their reason: a symbol's price history doesn't go back far enough for the window. Runs whose rule never fired would be kept too, drawn as hollow dots.

Explore and check. Plot any two of 13 metrics, filter by rule family, market, universe and window, and open any run to see its rule, its growth chart against buy-and-hold, the exact prompt and the exact engine requests behind it. Signed in, "Run this in the Strategy Agent" fills in the prompt for you. The analysis view gives each family's median Sharpe ratio and excess return with bootstrap intervals, and says those intervals understate the uncertainty because runs share windows and markets. Community templates now show their house backtest from the corpus, when the stored run used the template's current prompt.

Limits we state. The stock baskets are today's large companies, so their long windows carry survivorship bias; index funds don't have that problem. The Strategy Agent runs at most 10 years, so the 15 and 20-year runs call the engine directly, and the page says the agent prompt would run the last 10 years of the same rule.