AI-workflow for liquidation cascades in trading

Automated AI Workflow for Analyzing Liquidation Cascades in Trading

A liquidation cascade occurs when the price drops below a key level: the exchange closes positions with market orders, causing a sharp drop over 1-2 bars followed by a bounce. This pattern repeats dozens of times a year, but parameters change monthly—from support breakouts to spikes with 3% V-shaped reversals.

Manually updating criteria requires analyzing news about liquidated positions. An AI agent based on Claude with the /loop command automates the process, creating profitable Pine Script strategies for the current month.

To initiate: npx @backtest-kit/cli --init --output my-project.

System Prompt CLAUDE.md

The key element is the CLAUDE.md file with the agent's contract. It focuses on market thinking, not coding:

Mistakes in Strategy Development

• Avoid reading the entire project and cluttering context.
• Steer clear of brute force and incremental edits.
• Don't sacrifice efficiency for universality.
• Exclude var, na, side effects in Pine Script—calculate everything on each iteration.
• Prohibit hacks like infinite trailing SL without final criteria.
• Minimum 1 signal per day for statistical significance.

Correct Approach

• One-month strategy: ./math/jan_2026.pine, ./content/jan_2026.strategy.ts.
• Analyze negative news: "Bitcoin negative news March 2026 regulatory issues."
• Candle dump: npm start -- --dump --timeframe 15m --limit 500 --when "2026-02-28T00:00:00.000Z" --jsonl.
• Account for sideways markets, dynamic TP/SL (minimum TP 1% to cover 0.4% fees).
• No HOLD strategies—only entry points with immediate profit.

Agent Workflow Algorithm

Planning

• Analyze previous .pine and reasons for failure.
• Search for negative news from the month.
• Correlate news with candle dump (bounces, gaps, volatility).
• Review candles: volumes, market gaps, risks.

Writing and Validation

• New files from scratch, no copy-pasting.
• Run: npm start -- --pine ./math/impulse_trend_15m.pine --timeframe 15m --limit 500 --when "2026-02-28T00:00:00.000Z" --jsonl.
• Criterion: profit, not code for code's sake.
• Save report: ./report/feb_2026.md with fundamental analysis.
• Code review for HOLD, trailing SL, drawdowns.

Automation Example for February 2026

The agent received a task: January's strategy was unprofitable (bearish trend). A dump of 500 15m candles showed spikes and gaps. News: $700M liquidations on February 5-6, BTC drop to $60K.

The agent correlated events with data, generated a new .pine accounting for V-bounces, tested for profit >1% TP, >1 signal/day.

Results in Pine header:

• Sharpe Ratio, avgPnl, stdDev.
• Signals: 1+ per day.
• Risk management without HOLD.

Key Takeaways

• Monthly optimization based on news and candles eliminates whipsaw from fees.
• Dump and --jsonl for visual trend analysis (bullish/bearish/sideways).
• Negative news is key to correlating with liquidations.
• Code review is mandatory: final SL/TP, no brute force.
• Indicators only for predicting retail trader behavior.

— Editorial Team