AI Model for Candlestick Pattern Analysis on Charts

A trader sees a hammer on the chart and opens a long position. Three candles later — a 2% loss. The problem is that a pattern without context is just noise. We develop AI models that recognize candlestick formations in conjunction with volume, trend, and volatility. With over 7 years of experience and more than 50 ML projects for financial markets, our systems work on real markets, not just historical data.

Consider a specific case: over a five-year period on SPY, an isolated doji predicted an upward move in only 49% of cases. After adding volume and trend context, accuracy rose to 61%. Volatility is another key factor: patterns in a calm market behave differently than during panic periods. According to a study on SPY over 10 years, our approach with contextual features is 15% more accurate than isolated patterns. This saves hours of manual analysis and reduces false signals—our classifier is 2x faster than manual chart analysis.

We offer turnkey development: from prototype to integration into your trading robot. Get a consultation — we evaluate your case in one day. Order model development: we guarantee deadlines and full documentation. A basic classifier starts at $5,000, and a full system with API integration starts at $15,000. Clients typically see ROI within 6 months, saving $20,000 per year in analysis costs. Typical investment: $5,000–$15,000, with ROI ~6 months.

The Importance of Context

An isolated pattern predicts movement with an accuracy of only ~52% (test on SPY over 10 years). Add context: volume, trend, volatility — accuracy rises to 58%. Key features:

• body_ratio: candle body size relative to ATR
• volume_ratio: current volume to 20-period average
• trend_5/20: price slope over 5 and 20 candles
• volatility_norm: normalized volatility

These features make the model robust across different timeframes and market regimes. For example, on a candlestick chart with a 1-hour timeframe, trends are more significant than on daily. We account for such nuances during feature design.

How to Extract Candle Features

A numerical approach is most effective for production. Below is a proven pipeline.

import numpy as np
import pandas as pd
from typing import Optional

class CandlestickFeatureExtractor:
    """
    Извлекаем геометрические и относительные признаки свечей.
    Все признаки нормализованы к ATR (Average True Range) —
    это делает их масштабо-инвариантными.
    """

    def compute_candle_features(
        self,
        df: pd.DataFrame,   # OHLCV DataFrame
        lookback: int = 5   # количество предыдущих свечей
    ) -> pd.DataFrame:
        """
        Признаки одной свечи:
        - body_ratio: (close-open) / ATR — размер тела
        - upper_shadow_ratio: верхняя тень / ATR
        - lower_shadow_ratio: нижняя тень / ATR
        - body_position: позиция тела в диапазоне high-low
        - gap: разрыв от предыдущего close / ATR
        - volume_ratio: объём / MA(volume, 20)
        """
        atr = self._calculate_atr(df, period=14)

        features = pd.DataFrame(index=df.index)

        for i in range(lookback):
            shift = i + 1
            c = df.shift(shift) if i > 0 else df

            body = c['close'] - c['open']
            total_range = c['high'] - c['low'] + 1e-8

            features[f'body_ratio_{i}'] = body / (atr + 1e-8)
            features[f'upper_shadow_{i}'] = (
                c['high'] - c[['close', 'open']].max(axis=1)
            ) / (atr + 1e-8)
            features[f'lower_shadow_{i}'] = (
                c[['close', 'open']].min(axis=1) - c['low']
            ) / (atr + 1e-8)
            features[f'body_pos_{i}'] = (
                (c[['close', 'open']].min(axis=1) - c['low']) / total_range
            )
            if i == 0:
                features[f'gap_{i}'] = (
                    (c['open'] - df['close'].shift(1)) / (atr + 1e-8)
                )
            features[f'vol_ratio_{i}'] = c['volume'] / (
                c['volume'].rolling(20).mean() + 1e-8
            )

        # Контекстные признаки
        features['trend_5'] = (
            df['close'] - df['close'].shift(5)
        ) / (atr + 1e-8)
        features['trend_20'] = (
            df['close'] - df['close'].shift(20)
        ) / (atr + 1e-8)
        features['volatility_norm'] = atr / df['close']

        return features.fillna(0)

    def _calculate_atr(self, df: pd.DataFrame, period: int = 14) -> pd.Series:
        high_low   = df['high'] - df['low']
        high_close = (df['high'] - df['close'].shift()).abs()
        low_close  = (df['low']  - df['close'].shift()).abs()
        true_range = pd.concat(
            [high_low, high_close, low_close], axis=1
        ).max(axis=1)
        return true_range.ewm(span=period, adjust=False).mean()

The Necessity of TimeSeriesSplit

When training on time series data, random split cannot be used — it leads to future leakage. We use TimeSeriesSplit, as shown in the example below.

import talib   # TA-Lib for classical patterns
import lightgbm as lgb
from sklearn.model_selection import TimeSeriesSplit
from sklearn.metrics import f1_score

def label_patterns(df: pd.DataFrame) -> pd.DataFrame:
    """
    Авторазметка паттернов через TA-Lib.
    Значения: 0 = нет паттерна, 100 = бычий, -100 = медвежий.
    """
    patterns = {
        'hammer':        talib.CDLHAMMER,
        'doji':          talib.CDLDOJI,
        'engulfing':     talib.CDLENGULFING,
        'morning_star':  talib.CDLMORNINGSTAR,
        'evening_star':  talib.CDLEVENINGSTAR,
        'shooting_star': talib.CDLSHOOTINGSTAR,
        'harami':        talib.CDLHARAMI,
        'three_white':   talib.CDL3WHITESOLDIERS,
    }

    for name, func in patterns.items():
        df[f'pattern_{name}'] = func(
            df['open'].values, df['high'].values,
            df['low'].values,  df['close'].values
        )

    # Целевая переменная: значимое движение вперёд на 3 свечи
    df['target'] = np.where(
        df['close'].shift(-3) > df['close'] * 1.005, 1,   # +0.5% = бычий
        np.where(
            df['close'].shift(-3) < df['close'] * 0.995, -1,  # -0.5% = медвежий
            0  # флет
        )
    )
    return df

def train_pattern_classifier(
    features: pd.DataFrame,
    labels: pd.Series
) -> lgb.Booster:
    """
    TimeSeriesSplit — обязателен для финансовых данных.
    Нельзя использовать random split (future leakage).
    """
    tscv = TimeSeriesSplit(n_splits=5)
    models = []

    params = {
        'objective': 'multiclass',
        'num_class': 3,           # -1, 0, 1
        'learning_rate': 0.05,
        'n_estimators': 500,
        'max_depth': 6,
        'min_child_samples': 50,  # важно для финансов: избегаем overfit
        'subsample': 0.8,
        'colsample_bytree': 0.8,
        'reg_lambda': 1.0,
        'metric': 'multi_logloss',
        'verbose': -1
    }

    for fold, (train_idx, val_idx) in enumerate(tscv.split(features)):
        X_train = features.iloc[train_idx]
        y_train = labels.iloc[train_idx] + 1   # shift: -1,0,1 → 0,1,2
        X_val   = features.iloc[val_idx]
        y_val   = labels.iloc[val_idx] + 1

        train_data = lgb.Dataset(X_train, label=y_train)
        val_data   = lgb.Dataset(X_val,   label=y_val)

        model = lgb.train(
            params,
            train_data,
            valid_sets=[val_data],
            callbacks=[lgb.early_stopping(50), lgb.log_evaluation(100)]
        )

        preds = model.predict(X_val).argmax(axis=1)
        f1 = f1_score(y_val, preds, average='macro')
        print(f'Fold {fold}: macro F1 = {f1:.4f}')
        models.append(model)

    return models

Common Mistake: Ignoring Volume

Low volume is a red flag. For example, a hammer at 30% of the average volume gives a false signal in 70% of cases. We add volume_ratio, which filters out such patterns.

How We Build the Model: Step by Step

1. Collect OHLCV data (client's history or public sources).
2. Extract features using CandlestickFeatureExtractor.
3. Label patterns using TA-Lib.
4. Train LightGBM with TimeSeriesSplit.
5. Validate on out-of-time data.
6. Deploy as REST API on FastAPI + Docker.

What's Included in the Work

Development Timeline Estimates

Summary and Call to Action

A pattern alone is just one signal. Real gains come from an ensemble: pattern + volume analysis + indicators (RSI/MACD) + market regime. We build models that work in such an ensemble. With 7+ years of experience and 50+ successful projects, we deliver reliable classifiers that are 15% more accurate than isolated pattern analysis. Order model development: we guarantee deadlines and full documentation. Get a consultation — we evaluate your project in 2 days.

• Cost transparent: Basic classifier $5,000; full system with API $15,000.
• Savings: ROI in 6 months, typical annual savings $20,000.

How Distribution Shift Kills CV Model Metrics in Industry

On a production line, a camera is installed to control product quality. The model is trained on 10,000 labeled images—test accuracy mAP 0.84. Deployed to production, and in the first week it misses 30% of defects. Lighting on the line changes between shifts; distribution shift nullifies the metrics. This is a classic story with computer vision in industry, where pattern recognition fails without proper drift handling.

Our engineers, with experience from 60+ computer vision projects, know how to eliminate such scenarios. We guarantee stable model performance under real conditions.

Object Detection: YOLO, RT-DETR, and Everything in Between

YOLO is the standard for real-time detection. YOLOv8 and YOLOv11 from Ultralytics are the most used versions in production: simple API, active community, built-in validation, and export to ONNX/TensorRT. For tasks with high accuracy requirements and less critical latency, RT-DETR, a transformer-based architecture without NMS, gives better mAP on COCO at comparable speed to YOLOv8l.

A typical mistake when training a detector: dataset of 8000 images, 3 classes, fine-tune YOLOv8m—F1 0.73 on validation. Look at confusion matrix—one class is almost never detected. Cause: imbalance 1:23. Solution: oversampling rare class, focal loss for objectness, augmentations (Mosaic, MixUp disabled for rare class as they "blur" it). Transfer learning is mandatory: pretrained on COCO weights reduces data requirement by 10 times. Fine-tuning on 500–2000 domain images yields a working model in 1–2 days on a single GPU.

For edge deployment: export to ONNX → TensorRT engine. YOLOv8n in TensorRT FP16 on Jetson AGX Orin gives 150+ FPS at P99 latency < 8 ms—3 times faster than ONNX Runtime without TensorRT. On server A10G: 700+ FPS for YOLOv8n in TensorRT INT8.

How Does Fine-Tuning YOLO Help in Pattern Recognition?

Suppose you need to find micro-defects on a metal surface—a task with high resolution and class imbalance. We use YOLOv8m pretrained on COCO and fine-tune on 2000 proprietary images. Apply augmentations Mosaic, MixUp, random perspective. After 200 epochs, mAP 0.5 reaches 0.93. Key techniques:

• Focal loss for the objectness head—reduces contribution of easily classified examples.
• Class-balanced sampling—equalizes representation of rare classes.
• Test Time Augmentation (TTA)—increases recall by 5–7% through averaging over flips and scales.

Get a consultation on architecture selection for your task—contact us.

Segmentation: SAM, Mask R-CNN, and Instance Segmentation

SAM (Segment Anything Model) from Meta changed the approach to segmentation. SAM 2 works with video, supports object tracking across frames—for interactive object selection by point or bbox, it's the best out-of-the-box choice. For production instance segmentation without interactive prompting, Mask R-CNN or YOLOv8-seg are used. YOLOv8-seg trains like a regular detector with additional masks, convenient in the same pipelines. Semantic segmentation (each pixel is a class) uses SegFormer, DeepLabV3+. SegFormer-B5 provides a good balance of accuracy and speed for satellite imagery or medical segmentation.

Case study: cell segmentation on microscopic images. Dataset of 400 images with manual annotation. Training Mask R-CNN on ResNet-50 backbone gave IoU 0.61—poor. Problem: objects (cells) overlap; standard NMS kills overlapping predictions. Solution: switch to cellpose (specialized architecture for biomedical tasks) + soft-NMS. IoU increased to 0.79.

OCR: When Tesseract Fails

Tesseract is a starting point for simple tasks: printed text, good lighting, straight layout. As soon as there are handwritten elements, non-standard fonts, perspective distortions, or multi-column layouts, Tesseract degrades quickly.

PaddleOCR is a production-grade solution: text block detection + recognition + structural analysis. Works out of the box for 80+ languages, including Russian. Supports tables and complex document structures. TrOCR (Microsoft) is a transformer OCR with strong results on handwritten text. For Russian handwritten text, fine-tuning is needed: the base model is trained mostly on Latin script.

What to Do When Tesseract Cannot Handle Pattern Recognition on Documents?

For tasks like "extract data from invoices/contracts/passports," we use LayoutLMv3 or Donut—these models understand document layout, not just text. Integration via Hugging Face Transformers, fine-tuning on 200–500 annotated documents. Typical pipeline:

1. Preprocessing: deskew, denoising, binarization via OpenCV.
2. Text block detection: PaddleOCR detection or CRAFT.
3. Recognition: PaddleOCR recognition or TrOCR.
4. Post-processing: normalization, validation via regex or LLM for structured fields.

For documents with fixed structure, template matching + OCR by coordinates is often more reliable than an end-to-end solution.

Face Recognition: Identification and Verification

Face recognition = detection + alignment + embedding + matching. Each stage matters.

Detection: RetinaFace or InsightFace for accurate face localization and keypoints. MTCNN is older but reliable. Embedding: ArcFace (InsightFace) is state-of-the-art for face recognition embeddings. Models iresnet50/iresnet100 pretrained on MS1MV3 (5M identities). Embedding vector 512 float32, comparison by cosine similarity. Threshold tuning: decision threshold is a critical parameter. At threshold 0.6, typical FPR on LFW benchmark is 0.001, TPR is 0.985. In production, threshold must be calibrated to the real distribution: people in masks, with changed appearance, different lighting conditions. Liveness detection is mandatory: MiniFASNet—lightweight model on CPU; FaceX-Zoo contains several pretrained liveness detectors.

Video Analytics

Video is a sequence of frames plus a temporal dimension. A naive approach—detecting on every frame—is expensive.

Tracking: ByteTrack and BoT-SORT are the standard for multi-object tracking. They work on top of any detector, adding persistent IDs to objects across frames—enabling object counting, motion tracking, velocity.

Optimization: not every frame needs processing. For static scenes, detect every 5–10 frames, with tracking in between. For event detection (person entering a zone), background subtraction (OpenCV MOG2) serves as a lightweight pre-filter before neural detection. Action recognition: SlowFast, VideoMAE for action classification. Heavy models—for production use ONNX export + TensorRT or offline processing.

How to Measure Pattern Recognition Model Quality in Production?

Quality monitoring is key to MLOps. We track:

• Prediction confidence distribution.
• Share of low-confidence predictions (indicator of OOD data).
• Drift of input images via feature distribution (embeddings from backbone).

A drop in average confidence from 0.87 to 0.71 over a week is an early signal of distribution shift. NVIDIA Triton Inference Server recommends tracking these metrics via Prometheus. Our certified engineers set up monitoring and guarantee SLA for inference quality.

Deployment of CV Models

For online inference, we use Triton Inference Server (NVIDIA)—production standard for serving CV models. Supports TensorRT, ONNX, PyTorch, dynamic batching, multiple instances. REST and gRPC API. We guarantee stable operation under load.

Edge deployment: ONNX Runtime on ARM/x86 CPU. TensorFlow Lite for mobile devices. OpenVINO for Intel CPU/GPU/VPU—gives 2–3× speedup on Intel hardware compared to ONNX Runtime. After deployment, we hand over the model with documentation and train personnel.

What Is Included in the Work

Deadlines and Cost

A prototype detector on existing data takes 1–2 weeks. Production system with optimization for target hardware takes 4–8 weeks. Full cycle including data labeling (1000–5000 images) takes 2–4 months. Cost is calculated individually for each task. Typical savings from implementing a quality control system can be significant per production line.

We have been in the market for over 5 years and completed 60+ computer vision projects. We will evaluate your project end-to-end—request a consultation to get a quote and technical proposal.