Feature: Chart pattern similarity as observation/state for RL agents

[grahammccain]

Idea

FinRL agents currently use price/volume/technical indicators as state observations. Historical chart pattern similarity could add a useful signal — essentially asking "when the chart looked like this before, what happened next?"

Chart Library has 24M+ pre-computed pattern embeddings across 19K US equities (10 years). The API returns the top-K most similar historical patterns and their forward returns, which could serve as additional state features for RL agents.

How it could work in FinRL

import requests

def get_pattern_features(symbol: str, date: str) -> dict:
    """Add pattern similarity features to FinRL state space."""
    resp = requests.get("https://chartlibrary.io/api/v1/search", params={
        "symbol": symbol, "date": date, "timeframe": "RTH"
    }, headers={"X-API-Key": "your-key"})
    
    matches = resp.json()["matches"]
    
    # Extract features for RL state
    avg_1d_return = sum(m["return_1d"] for m in matches) / len(matches)
    avg_5d_return = sum(m["return_5d"] for m in matches) / len(matches)
    avg_10d_return = sum(m["return_10d"] for m in matches) / len(matches)
    win_rate = sum(1 for m in matches if m["return_5d"] > 0) / len(matches)
    avg_distance = sum(m["distance"] for m in matches) / len(matches)
    
    return {
        "pattern_avg_1d": avg_1d_return,
        "pattern_avg_5d": avg_5d_return,
        "pattern_avg_10d": avg_10d_return,
        "pattern_win_rate": win_rate,
        "pattern_confidence": 1.0 / (1.0 + avg_distance),  # closer = more confident
    }

# These features could be added to StockTradingEnv observation space

Why this might help

• Pattern similarity captures non-linear chart structure that moving averages and RSI miss
• The forward returns from historical matches act as a "base rate" prior
• Low-distance matches (high confidence) correlate with more predictable outcomes
• Could be especially useful for the regime-switching aspects of trading

Practical details

• Free tier: 200 calls/day, enough for daily rebalancing experiments
• Response time: ~100ms per search
• Docs: https://chartlibrary.io/developers
• Coverage: 19K US equities, 8 timeframes, 10 years

Curious if the team has considered pattern-based features in the observation space. Happy to discuss.

[atharvajoshi01]

Interesting idea but the tight coupling to a commercial API with a 200 calls/day free tier is a concern for an open source research framework. A 2-year daily backtest on 10 stocks needs ~5k API calls, which exceeds the free tier. Also hard to guarantee no look-ahead bias if the embedding database overlaps with the backtest period.

A more self-contained approach would be implementing pattern matching locally using DTW or learned embeddings, then exposing a generic interface where Chart Library could be one optional backend. That way the core feature stays reproducible without an external dependency.