AI Strategy Optimizer

addOptimizerSchema takes the training/test ranges, the source list, and a getPrompt that turns the assembled multi-timeframe messages into one strategy report; Optimizer.dump runs it and writes the executable:

import { addOptimizerSchema, Optimizer, listenOptimizerProgress } from "@backtest-kit/ollama";

addOptimizerSchema({
  optimizerName: "btc-optimizer",
  rangeTrain: TRAIN_RANGE,   // 7 daily ranges, Nov 24–30 2025
  rangeTest:  TEST_RANGE,    // Dec 1 2025
  source:     SOURCE_LIST,   // 4 timeframe sources (below)
  getPrompt: async (symbol, messages) => text(symbol, messages),  // LLM → strategy report
});

listenOptimizerProgress(({ progress }) => console.log(`Progress: ${progress * 100}%`));

await Optimizer.dump("BTCUSDT", { optimizerName: "btc-optimizer" }, "./generated");
Copy

Optimizer.getData("BTCUSDT", { optimizerName }) is available too — it returns the per-range strategy metadata without generating code (commented out in the demo, handy for inspecting what the LLM produced).

Each source is { name, fetch, user, assistant }: fetch pulls rows from CCXT_DUMPER_URL/view/<range> for the given symbol/date window; user renders them as a markdown table followed by an exhaustive indicator legend (period, lookback, range for every column); assistant is the acknowledgement turn.

const SOURCE_LIST = [
  { name: "long-term-range",  /* 1h, 48 candles, RSI/MACD/ADX/Bollinger/Fibonacci/DEMA/WMA/SMA50 … */ },
  { name: "swing-term-range", /* 30m, 96 candles, + volume/volatility/Bollinger width */ },
  { name: "short-term-range", /* 15m, fast indicators + ROC(5/10) */ },
  { name: "micro-term-range", /* 1m, 60 candles, squeeze momentum + pressure index */ },
];
// each: fetch({symbol,startDate,endDate,limit,offset}) → rows; user(symbol,data) → markdown + legend
Copy

The legends are deliberately verbose — they tell the model exactly what each column means (e.g. "RSI(14): over previous 14 candles before row timestamp, Min 0 Max 100"), so the generated strategy reasons over named indicators rather than raw numbers.

text() calls Ollama with think: true, a system instruction to return only a copy-paste-ready strategy report (no greeting, no meta-talk), the assembled source messages, and a final question asking for a fundamental (not just technical) recommendation with S/R entries and an RR target. The returned report is escaped for safe embedding into the generated .mjs:

const response = await ollama.chat({
  model: "deepseek-v3.1:671b", think: true,
  messages: [
    { role: "system", content: "только отчёт готовый для копипасты … Не здоровайся, не говори что делаешь" },
    { role: "system", content: "Reasoning: high" },
    ...messages,
    { role: "user", content: `На каких условиях мне купить ${symbol}? … RR? LONG или SHORT? … фундаментальный анализ` },
  ],
});
Copy

./generated/btc-optimizer_BTCUSDT.mjs is self-contained: exchange config (Binance/CCXT), frame definitions (the 7 training days + the test day), 7 strategy implementations (one per training day's generated report), a Walker setup, and progress listeners. Running it executes all seven strategies on the test data and ranks them by Sharpe — so a week of LLM analysis collapses into one comparable, reproducible artifact.