Prophecy Trader — a self-learning stock and crypto paper-trading bot that never forgets

Train it in 4 steps

1. Load data — start with a DEMO, then load a real stock (Stooq) or coin (CoinGecko), or upload a Yahoo Finance CSV.
2. Train — press Train. A population of strategies competes; winners breed. Watch fitness climb and the test return (unseen data) — that is the honest number.
3. Check the memory — the Lessons tab shows which market patterns paid. Use 👍 / 👎 to reinforce or punish a pattern yourself.
4. Paper-trade — Step 1 bar to replay history, or start paper autopilot once readiness ≥ 60. Every fill is simulated with commission + slippage.

How it learns (like Evolution Arena)

Each strategy is a genome: weights on trend, momentum, RSI, MACD, Bollinger, volume and volatility, plus stop-loss, take-profit and position size. Fitness is measured by backtesting on the train slice; the all-time champion, the hall of fame and every pattern lesson persist in your browser forever — that is the “never forgets” memory. Export the brain JSON to back it up or move it to another device.

Broker bridge — real paper orders via Alpaca

1. Sign up free at alpaca.markets (paper trading is free worldwide, no brokerage account needed).
2. In the dashboard, switch to your Paper Trading account and generate paper API keys.
3. Paste the key ID + secret into section 5 above and press Connect.
4. Load a live US stock (e.g. AAPL) or BTC/ETH, train until readiness ≥ 60, tick both routing boxes, start autopilot.
5. Signals now execute as genuine paper orders through Alpaca's systems — fills appear in the Journal tagged ALPACA.

Why not PayPal? PayPal is a payments app, not a broker: it offers no stock trading and no API for bots to trade with. Some brokers accept PayPal for deposits, but trading itself always happens through a broker's API — which is what this bridge uses, on the paper endpoint only.

Could it trade real money one day?

Honestly: this page alone never will, by design. Real automation would need a broker with an API, a server to hold the keys (never a browser page), and careful handling of UK tax and the FCA perimeter around investment activity. The realistic path is: months of profitable paper trading here → tiny-size live testing via a regulated broker's own tools → only then consider automation, ideally with professional advice. Anyone promising guaranteed bot profits is selling something.

Reading the Reality Check panel (and the Stress test tab)

Most of the time a "great" backtest is an accident. The new Reality Check panel shows a small battery of independent tests on the champion's out-of-sample slice:

• Train/test gap — the bigger the gap, the more the strategy has memorised its training window instead of learning a real pattern. Above 30% is a strong overfit warning.
• Calmar & Sortino — return per unit of pain. Calmar uses the worst peak-to-trough drop; Sortino only counts the losing days. Both above ~1 is decent; above 2 is rare.
• Max losing streak — how many losing trades in a row the strategy is expected to hit. Strategies with long streaks need emotional stamina you may not have.
• % winning months — even great strategies lose months. Below 50% is a warning sign.
• Significance (p) — a permutation test: how often would random shuffling of the trade list match or beat your result? p < 0.10 is "maybe real", p < 0.01 is "probably real".
• After fees & tax — headline return minus a coarse model of UK stamp duty (UK shares), spread, and 18% capital gains tax. This is the number that actually ends up in your pocket.

The Stress test tab adds three more checks. Anchored OOS re-runs the champion on data the GA has never trained on (a held-out slice from the very start); passing ≥ 2 of 3 windows is a hopeful sign. Cross-symbol sanity runs the same champion genome on entirely different synthetic series — a strategy that's only profitable on the original symbol is almost certainly overfit. Monte-Carlo reshuffles your trade order 200 times to estimate the worst drawdown you should expect even if the average is fine — the 95% number is the one to size your position around.

The Brain tab now explains what the champion actually does

Each champion is just a vector of numbers — feature weights, thresholds, position size. That's hard to interpret. The Brain tab now shows a plain-English summary that turns the genome into a sentence: "Entry: looks for trend + momentum + RSI to align, then enters moderately. Risk: moderate stops at 4%, small targets at 6%. Sizing: medium positions." Use it to sanity-check that the bot is doing something a human would actually do, not something only an optimisation algorithm could love.

What changed in the fitness function

The earlier fitness function rewarded raw annualised return very heavily (×200 in "profit" mode). That's exactly the knob that makes a GA find strategies that win big once and lose steadily. The new fitness treats return and max drawdown on equal footing, with a separate penalty for downside variance of per-trade returns. Practically: a strategy that makes 50% with a 60% maxDD is now scored worse than a strategy that makes 25% with a 15% maxDD, even though the first is "more impressive" in a screenshot.

• Overfit to the training slice. The genetic algorithm is brilliant at finding rules that perfectly describe the past and perfectly fail the future. This is the #1 reason "I made 40% in backtest!" turns into "I lost 30% in a week" live.
• Survivorship in the symbol list. Only testing stocks that exist today — the ones that went to zero were quietly dropped. The bot only sees data you feed it.
• Ignoring costs. Commission + slippage + spread + stamp duty (UK 0.5%) + CGT turn "amazing 60% annual return" into "small positive return" very quickly on a high-turnover strategy. The Reality Check shows this.
• Not enough trades. A strategy with 8 trades in the test slice has no statistical meaning. We require ≥ 10 before autopilot unlocks.
• Curve-fitting the random seed. Train twice with different starting populations and you'll get two "amazing" champions. If neither survives Walk-Forward, the edge was the noise, not the signal.
• Regime change. A strategy that worked in 2020's crash may not work in 2024's grind. Even good strategies stop working. The honest answer is always: re-test on the most recent data you have.

Tools · Music · Donate · Legal · Help & Contact · Data: Stooq (stocks, free) & CoinGecko (crypto, free). Uploaded files never leave your device.