Chart pattern recognition system development

Chart Pattern Recognition System Development (Head and Shoulders, Triangles, Wedges)

Automatic recognition of classical chart patterns is a task that at first glance looks simple, but in practice requires serious engineering work. The "head and shoulders" pattern is easy to see on a static chart, but detecting it algorithmically in real-time across thousands of instruments is a different story.

Pattern Classification and Algorithmic Representation

All patterns can be divided into three categories by detection complexity:

Reversal patterns (Head & Shoulders, Inverse H&S, Double Top/Bottom, Triple Top/Bottom) — formed at significant levels, signal trend reversals.

Continuation patterns (Ascending/Descending/Symmetrical Triangle, Bull/Bear Flag, Pennant, Wedge) — formed in the middle of a trend as consolidation before continuation.

Channel patterns (Rising/Falling Channel, Rectangle) — price moves between parallel levels.

Detection Algorithm: Pivot Points Method

The main approach is identifying local extremes (swing highs and swing lows) and analyzing their sequence.

Stage 1: Finding pivot points

def find_pivots(highs, lows, window=5):
    pivot_highs = []
    pivot_lows = []
    for i in range(window, len(highs) - window):
        if highs[i] == max(highs[i-window:i+window+1]):
            pivot_highs.append((i, highs[i]))
        if lows[i] == min(lows[i-window:i+window+1]):
            pivot_lows.append((i, lows[i]))
    return pivot_highs, pivot_lows

Window size affects the "scale" of detected patterns. Small window gives patterns on short timeframes, large — on long-term.

Stage 2: Pattern Matching

For "head and shoulders" we need to find a sequence of 5 pivot points: left shoulder (high), neckline left (low), head (highest high), neckline right (low), right shoulder (high).

Conditions for valid H&S:

• Head higher than both shoulders (tolerance ±2%)
• Both shoulders approximately same height (difference < 5%)
• Neckline relatively horizontal (slope < 15°)
• Right shoulder does not exceed Head

Stage 3: Triangle Detection

Triangles are defined through linear regression on pivot points:

• Symmetrical: descending line on highs + ascending on lows, converging to one point
• Ascending: horizontal resistance + ascending support
• Descending: horizontal support + descending resistance

Triangle quality is assessed through R² coefficient of linear regressions. R² > 0.85 — good triangle.

Stage 4: Wedges

Wedge differs from triangle in that both lines point in one direction:

• Rising Wedge (bearish signal): both lines ascending, upper more gradual
• Falling Wedge (bullish signal): both lines descending, lower more gradual

Key condition: lines converge (angle between them decreases).

Pattern Quality Assessment and Target Projection

Each detected pattern receives a score (0–100) based on:

• Symmetry (for H&S)
• R² of lines (for triangles)
• Volume ratio (volume should decrease during formation)
• Pattern completeness (has breakout occurred or not)

Target projection — calculation of target after breakout:

• H&S: target = neckline level − head height
• Triangle: target = apex point ± base height
• Wedge: target = wedge start

Backtesting and Reliability Assessment

System includes backtesting module: for each historically detected pattern we check if target worked within the next N candles and with what success rate. This allows calibrating threshold values for specific asset and timeframe.

Typical reliability in crypto market (based on historical BTC/ETH data):

System Architecture

Backend: Python (pandas, numpy, scipy for linear regression), OHLCV data processing from exchange APIs (CCXT library). Scan across all instruments is scheduled (cron) on each candle close.

Database: PostgreSQL for storing detected patterns with parameters, status (forming/confirmed/failed) and results.

Frontend: React + TradingView Lightweight Charts. Patterns are drawn as SVG overlays on the price chart — shoulder lines, neckline, triangle lines with labels.

Alerts: When a new confirmed pattern is detected or a level is broken — notification via Telegram or webhook.

Scaling

For scanning 500+ instruments in real-time, parallel processing is used through Python multiprocessing or Celery. Results are cached in Redis. Full scan of 500 instruments on 4h timeframe takes 15–30 seconds on standard server.

We develop a system with configurable detector sensitivity, support for multiple timeframes simultaneously, pattern visualization on chart and built-in backtesting for assessing algorithm quality on historical data.