Crypto markets aren't first-order chaotic anymore. Thousands of identical RSI / MACD / Bollinger bots execute the same signal in the same 100 ms window — they create the noise they're trying to trade. The only beneficiary is the exchange's fee meter.
Strategy alpha decays. The runtime underneath it doesn't. We invest where most desks under-invest — in the boring infrastructure that makes every iteration faithful, every signal reproducible, every deploy painless.
Every data fetch resolves against an AsyncLocalStorage temporal context. Your strategy can't peek at future bars even if you try — the runtime refuses to surface them. Multi-timeframe joins synchronize automatically.
// no timestamp params. context flows. const c1h = await getCandles(s, "1h", 100); const c15m = await getCandles(s, "15m", 100); const c5m = await getCandles(s, "5m", 100); // → all three resolve to the // same backtest tick
Eight specialised agents hit Tavily and the open web for acute event triggers — SEC actions, exchange hacks, presidential statements — within the last 4–12 hours. Vector search retrieves narrative; the LLM decides BUY / SELL / WAIT with cited reasoning.
await search("BTC breaking news ${date}");
await search("SEC CFTC DOJ action ${date}");
await search("Trump statement BTC ${date}");
await search("flash crash reason ${date}");
// → typed signal + reasoning + cited
// sources, validated by ZodPineTS executes .pine files locally on any data source. Strategies built for BTC on Binance redeploy unchanged to UZSE (Uzbekistan), MSE (Mongolia), DSE (Dhaka) — markets TradingView refuses to list.
npm start -- \ --pine ./math/feb_2026.pine \ --timeframe 15m --limit 500 \ --when "2026-02-28T00:00:00Z" \ --jsonl // → dump consumable by any LLM // or downstream notebook
Strategies, risk rules and frame definitions are first-class entities in a typed swarm. The same monorepo runs hundreds of strategies in parallel against shared exchanges and shared risk validators — no copy-paste, no drift.
addStrategySchema({
strategyName: "liq_spike_bounce",
interval: "15m",
riskList: [
RiskName.TakeProfitDistanceRisk,
RiskName.StopLossDistanceRisk,
],
getSignal,
});Output from the live ReAct pipeline — search the news, weigh contradictions, emit one signal with a written justification. Pick a date to see what the agent saw, queried, and decided.
The acute event driver — Trump's announcement of a two-week ceasefire with Iran (04:30 UTC) — triggered an instant rally to $72,000, liquidation of ~$425M in short positions, and a massive volume spike (>2M BTC/hour). Facts indicate a sharp bullish impulse that should sustain price growth in the coming hours. Optimal move: open a long with a tight stop-loss around $70,000.
A Telegram signal channel claims alpha on TRXUSDT. We treat it like an OSINT target — fingerprint the bot, geolocate the manipulation window, then trade the inversion. The seller's edge becomes ours.
| # | Entry | TP1 | TP2 | TP3 | TP4 |
|---|---|---|---|---|---|
| 1 | $0.0712 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 2 | $0.0698 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 3 | $0.0721 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 4 | $0.0688 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 5 | $0.0702 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 6 | $0.0731 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 7 | $0.0695 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
| 8 | $0.0709 | ×1.52 | ×1.74 | ×1.47 | ×1.50 |
Eight SHORT calls at the bottom of the 4H candle — every one moved against the recommendation within 45 minutes. Net of fees and slippage, the inverted book closed January 2026 at +8.54% with a portfolio Sharpe of 1.08.
Monte-Carlo across our published backtest distributions. Adjust capital, risk-per-trade and horizon — equity, drawdown and trade count update in real time. Production runs the same code paths against live exchange feeds.
From a single discovery call to multi-year asset mandates. Every engagement runs on the same open-source toolchain — so the strategy you commission today is auditable, reproducible, and yours to keep running.
We don't sell a black box. The toolchain that powers our strategies is published on GitHub, structured as a monorepo for parallel execution. Fork it, audit it, run it on your laptop.
Look-ahead bias, missing stop-loss, invalid risk-reward — all caught by Zod validators and AsyncLocalStorage guards before a signal leaves the boundary.
Hundreds of strategies share one exchange handle, one risk validator pool, one queue. Read the architecture →
Every tick, every reasoning chain, every risk rejection — appended as JSON Lines. Claude Code reads them natively; jq them in a second.
AsyncLocalStorage as a temporal context propagated through every async boundary. Strategies physically cannot peek at future data.
Why 1,000 identical RSI bots lose 18% / day to each other — and what to use instead of canned indicator stacks.
Claude Code finally has a TradingView replacement — local Pine Script, JSONL logs, an agent skill that survives context resets.
Gamma convexity, jump diffusion, and the 2-and-20 incentive structure behind liquidation cascades.
Claude's /loop command, calendar-month strategy files, and an acceptance contract that prevents brute-force iteration.
Why TauricResearch and node-ccxt-backtest both fail — and what changes when the agent searches before it decides.
Vector search over near-zero-score results. Why a 24-hour window beats a 4-hour one for finding regime change.
Inverting a Telegram signal channel's TRXUSDT calls. Sharpe 1.08, PnL +8.54%, 8/8 inversions correct.
UZSE, MSE, DSE — eight regional exchanges with no TradingView listing. Scrape, synthesize candles, run Pine.
Where +20%/month DCA returns actually come from — and what brokers hide on the dashboard while you wait for them.
npx -y @backtest-kit/cli --init scaffolds a working strategy in 30 seconds. When you're ready for a custom mandate, we're at tripolskypetr@gmail.com.