A genetic algorithm breeds grid-DCA trading strategies on real BTC candles, kills everything that fails on data it has never seen, and paper-trades the survivor. Live, in your browser.
Run it locally → · How it works · Honest results · Promo video
96 configs per generation · 8 genes · 4 species · 2 399 real BTCUSDT hours · No build, no npm install, no API keys
Every 4.8 seconds a new generation is born. Immigrants and offspring appear in the gene pool, get backtested, face the out-of-sample gate, and everything that is not an elite dies. The best config that survives the gate takes over the paper grid.
| Evolve | Select | Trade |
|---|---|---|
| Crossover and mutation over 8 genes. Species quotas stop one lucky family from wiping out the rest. | Train on 70% of the tape, then gate on the unseen 30%. Only configs that stay profitable with low drawdown survive. | The leader runs a real grid-DCA bot on the out-of-sample candles: fills, take-profit, stop, fees. |
- Run / Pause, Step to the next generation, and 1× 2× 4× 8× speed.
- Click any node in the gene pool to inspect its genome, train and out-of-sample results. Click empty space or press Esc to return to the leader.
- Change the seed to grow a completely different evolution. Same seed, same result, every time.
- Keyboard: Space pause, → step, 1–4 speed.
- URL options:
?seed=42,?speed=4,?warm=50(evolve 50 generations instantly on load),?paused. - Works on phones: panels stack, nothing scrolls sideways.
Full-page screenshot: docs/images/dashboard.png
25 s promo video · click to play
Requires Python 3 (to serve the files) and any modern browser.
git clone https://github.com/YOUR_GITHUB_USERNAME/sets-machine.git
cd sets-machine
python -m http.server 8000 --directory distOpen http://localhost:8000. That is all: no npm install, no build step, no keys, no database. Everything runs client-side in plain ES modules.
Host it for free on GitHub Pages: Settings → Pages → Deploy from a branch → main / (root). The root index.html forwards visitors to dist/.
Run the tests (Node 18+, no dependencies):
node --test tests/*.test.mjsRefresh the market data from Binance's public API (standard library only, no key):
python tools/fetch_data.py --hours 2400Each strategy is a long-only grid-DCA bot described by 8 genes:
| Gene | Range | What it does |
|---|---|---|
family |
4 species | Entry logic: Momentum (z-score above +Z), Mean revert (below −Z), Vol breakout (close above the previous N-bar high), Range grid (inside a quiet band) |
lookback |
10–200 h | Window for the rolling mean, deviation and breakout high |
entryZ |
0.2–2.5σ | How far price must stretch before the bot enters |
levels |
2–8 | Number of buy orders in the grid |
spacing |
0.3–3% | Distance between grid levels |
mult |
1–2× | Size multiplier per deeper level |
tp |
0.3–4% | Take-profit above the average entry |
stop |
1–12% | Stop below the deepest level; closes everything |
| Stage | In the code |
|---|---|
| Observe | Read volatility of the current tape window |
| Hypothesize | Inject 8 random immigrants |
| Mutate | Tournament selection inside each species, uniform crossover, Gaussian mutation (p = 0.18) |
| Backtest | Every newcomer is backtested on train (70%) and out-of-sample (30%) |
| Select | Keep the top 4 overall plus the best of each species; everyone else dies |
| Deploy | Best train fitness among configs that pass the gate becomes the paper-trading leader |
Fitness: train return − 0.6 × max drawdown, with a penalty below 4 trades.
Gate: out-of-sample return > 1%, drawdown < 10%, at least 3 trades, win rate ≥ 50%.
- Signals are computed on the close of bar j and executed at the open of bar j + 1. A test proves that changing a future candle cannot change a past signal.
- Inside a bar, fills are processed adverse-first: grid fills, then stop, then take-profit.
- Every fill and exit pays a 0.04% fee. Open positions are marked out at the end of a test window.
- The paper-trading panel replays the out-of-sample candles with the same
GridBotclass that the backtests use, so what you see is what was scored.
A sample of runs after 50 generations on the bundled data (train 2026-06-23 → 08-23, out-of-sample 08-23 → 09-22):
| Seed | Leader species | Train return / DD | Out-of-sample return / DD | OOS trades | Buy & hold (same 30 days) |
|---|---|---|---|---|---|
| 2026 | Range grid | +28.8% / 4.5% | +8.4% / 5.3% | 5 | +11.4% / 7.7% |
| 7 | Vol breakout | +28.6% / 4.3% | +8.3% / 7.7% | 4 | +11.4% / 7.7% |
| 42 | Mean revert | +30.8% / 2.9% | +7.3% / 3.9% | 54 | +11.4% / 7.7% |
Read this before getting excited:
- On this window buy & hold made more money. The evolved grids made less, with a smaller drawdown in two of three runs.
- The out-of-sample window is 30 days with a handful of trades. That is a sanity check, not proof of an edge.
- Picking the leader from configs that passed the gate reuses the out-of-sample data, so its numbers are optimistic.
- High win rates come from wide stops that were never hit in this window. That is exactly the risk a grid carries.
SETS MACHINE is a transparent research toy for watching evolutionary search work. It is not a trading bot to connect to real money.
index.html Redirects to dist/ (for GitHub Pages)
dist/
index.html Dashboard shell and controls
app.js Controller: generation timeline, paper trading, UI state
style.css Responsive blue-and-white interface
engine/
series.js Causal rolling mean / deviation / breakout high
bot.js GridBot: signals, grid fills, TP, stop, fees, metrics
evolution.js Genome, crossover, mutation, species quotas, gate
rng.js Seeded randomness
ui/draw.js Canvas painters: logo, ring, fitness, gene pool, Kelly, chart
data/candles.js 2 399 hourly BTCUSDT candles from Binance
assets/ Fonts, logo, favicon
tools/fetch_data.py Refreshes dist/data/candles.js from Binance's public API
tests/ Node test runner: data, indicators, bot mechanics, GA
docs/ README images, GIFs and the promo video
Built with plain HTML, CSS and JavaScript modules on <canvas>. No frameworks and no external requests at runtime.
- Market data: Binance public market data API, BTCUSDT 1h klines.
- Fonts: JetBrains Mono and Space Grotesk, both under the SIL Open Font License 1.1.
- Kelly criterion: J. L. Kelly Jr., A New Interpretation of Information Rate (1956).
Code is MIT licensed, see LICENSE. Test results are in VALIDATION.md.
Not financial advice. Paper trading on historical data only. No exchange connection, no keys, no real orders. Past performance, simulated or not, does not predict future results.