Start here
A paper-money experiment in evolving trading strategies and testing them honestly.
What this is
Evolution Trader is an experiment. It breeds trading strategies on a computer, the way breeders select animals, and then tests them as honestly as it can. Every strategy trades paper money only: no real money is ever involved. The hard part is not finding a strategy that looked great in the past (that is easy, and mostly luck). It is finding out whether any of it holds up on days the strategy never saw.
The honest answer so far
Read these as leads, never as wins. A single run can look great by luck, and only the forward test cannot be fitted to the past.
- S&P 500 (run of 2026-10-14; engine: current engine; same result in 3 reruns): the champion made +7.5% against SPY +10.3%. Luck check: it beat 58% of fair random portfolios (50 is average luck); textbook momentum made +8.6%; window 2026-04-16 to 2026-10-14, a period this run's training never saw. Its two biggest stocks made 42% of the gain (SNDK INTC). Across 2 distinct current-engine runs (their windows differ) the champion's holdout return ranged from +0.5% to +7.5%, so one run is a lead to test, not a result. 2 earlier runs (before 2026-10-15 15:00 UTC) came from an older engine on different windows, so they are not compared here.
- Nasdaq-100 (run of 2026-10-14; engine: current engine): the champion made +99.8% against QQQ +14.3%. Luck check: it beat 100% of fair random portfolios (50 is average luck); textbook momentum made +32.0%; window 2026-04-16 to 2026-10-14, a period this run's training never saw. Its two biggest stocks made 38% of the gain (MU MRNA). One run is a lead to test, not a result. 2 earlier runs (before 2026-10-15 15:00 UTC) came from an older engine on different windows, so they are not compared here.
- Stock-picking signals (S&P 500, 101 months): |t| < 2 for every signal: no signal reliably predicted next month.
- Forward test: not yet, no days recorded. It starts 2026-10-16.
Forward-test days
0
Not yet started
Forecasts resolved
18
Our Brier 0.197 vs coin flip 0.250 (lower is better); fewer than 30, so too few to judge
Evolution runs on record
9
Distinct results across three arenas (11 rows counting reruns), plus 20 walk-forward window rows
Where to go next
Forward testLive paper results that cannot be fitted to the past.Evolution runsEach run's best strategy against SPY/QQQ and textbook momentum.The storyEight slides: the question, the traps and what we found.How it worksThe pipeline, and the mistake each stage prevents.Reading the reportsWhat each Discord post and file means.GlossaryPlain-English definitions of every term.ForecastingHow well our probability forecasts match what happened.What we've learnedFindings, experiments and the decision log.CEO digestsThe latest plain-English status report.