AI Order Flow Analysis Model

You trade E-mini S&P 500 futures. The price stays in a narrow range for 15 minutes, but the order book volume grows. Regular indicators say nothing, while the Order Flow model sees: buyer volume aggressively exceeds seller volume, CVD is rising, the Footprint shows clusters at the 4500 level. Two minutes later, the price breaks the level from below — you know in advance it is a false breakout. Our AI Order Flow analysis model provides this advantage. Solutions are already deployed in 10+ projects, and our team has over 5 years of experience in AI/ML for financial markets.

How Does Aggressor Classification Work?

Every trade has an initiator — the aggressive side. According to the Lee-Ready algorithm, proposed over 30 years ago: a trade at a price higher than the previous is buyer-initiated, lower is seller-initiated. If the price is unchanged, we look at the previous movement. We fine-tune the classifier on labeled data, achieving up to 95% accuracy for cryptocurrencies (Binance) and 92% for futures (CME).

What Are Delta and CVD?

CVD (Cumulative Volume Delta) is a key Order Flow indicator:

Delta = Buyer_Volume - Seller_Volume
CVD = Σ Delta over period

Positive CVD with rising price = trend confirmation. Negative CVD with rising price = divergence, often preceding a reversal. We build CVD for multiple time windows (1s, 30s, 5min) and feed it as a feature into the model. Additionally, Absorption: when a large player absorbs aggressive orders without price movement. These are support/resistance levels that the model learns to detect.

Feature Engineering: From Ticks to Features

Tick data is transformed into features via rolling window aggregation:

def compute_order_flow_features(trades_df, window_seconds=60):
    features = {}
    trades_df['initiator'] = np.where(trades_df['side'] == 'buy', 1, -1)
    features['buy_volume'] = trades_df[trades_df.initiator==1]['volume'].rolling(f'{window_seconds}s').sum()
    features['sell_volume'] = trades_df[trades_df.initiator==-1]['volume'].rolling(f'{window_seconds}s').sum()
    features['cvd'] = features['buy_volume'] - features['sell_volume']
    features['trade_imbalance'] = features['cvd'] / (features['buy_volume'] + features['sell_volume'])
    features['avg_buy_size'] = (features['buy_volume'] / buy_count)
    features['avg_sell_size'] = (features['sell_volume'] / sell_count)
    features['large_buy_ratio'] = (large_buy_volume / total_volume)
    return features

Volume Profile — a histogram of volume at price levels. VPOC (Volume Point of Control) — level with maximum volume, used as support/resistance. Time and Sales analysis: clusters of large trades in a short time = large player entering a position.

Footprint CNN: When a Neural Network Reads Clusters

A Footprint Chart (Cluster Chart) combines Order Book and Order Flow: each candle is split into price levels, each level contains [buyer_volume × seller_volume]. We feed this data as a 3D tensor [time_bins × price_levels × 2] into a convolutional network:

class FootprintCNN(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv3d(1, 32, kernel_size=(3, 3, 2))
        self.conv2 = nn.Conv3d(32, 64, kernel_size=(3, 3, 1))
        self.flatten = nn.Flatten()
        self.fc = nn.Linear(64 * ..., 1)

The CNN learns to detect divergences and absorption that are inaccessible to linear models.

Comparison of Approaches: ML vs Classic Indicators

The CNN on footprint data outperforms regular technical indicators by 2-3 times in price direction prediction accuracy on short intervals.

Accuracy Comparison by Instrument (1-minute forecast)

The CNN consistently outperforms classic ML models on all instruments, especially on volatile markets.

AI Model Deployment Process

1. Data audit — assess tick data quality, select source, calculate required volume (minimum 3 months of history).
2. Feature Engineering — develop Order Flow, Volume Profile, Stacked Imbalance features.
3. Baseline model — Linear Regression or LightGBM for a quick baseline (3-4 weeks).
4. Advanced model — Footprint CNN with backtesting and optimization (3-4 months).
5. MLOps Pipeline — deploy model to production (Kubernetes, vLLM, Kafka for streaming).
6. Monitoring and drift — automatic quality reassessment, alerting on metric drops.

What Is Included in the Result

• Source code of the model and pipelines (Python, PyTorch/TensorFlow).
• Documentation: architecture description, run instructions, API specification.
• Team training (2-3 sessions of 4 hours each).
• Technical support for 3 months after deployment.
• Guarantee: if the model does not meet target metrics (ROC-AUC ≥ 0.7 on validation), we refine it free of charge.

Estimated Timelines

• Order Flow Feature Engineering + baseline regression: 3-4 weeks.
• Footprint CNN with backtesting and production pipeline: 3-4 months.
• The cost is calculated individually after a data audit and defined KPIs. Contact us for a consultation — we will select the optimal solution for your infrastructure and budget. Order a pilot project: we will conduct a data audit and show a baseline model in 2 weeks.

When does a time series forecasting model fail in production?

The CFO requests a quarterly sales forecast. An analyst builds SARIMA on three years of data, achieves MAPE 8.3% on the test set, and deploys. Two months later, the metric in production jumps to 23%. The root cause: the model was trained on pre‑COVID data, tested on a stable period, but production hit a promotion and supply chain disruption. Data leakage plus distribution shift—perfect notebook numbers, a broken forecast in reality. We have seen this pattern dozens of times across retail, fintech, and IoT. Our team has delivered more than 50 forecasting projects over 5+ years.

Incorrect cross-validation. Standard train_test_split for time series creates data leakage: the model sees future values during training. The correct approach is TimeSeriesSplit or walk‑forward validation with an expanding window.

Multiple seasonality. Hourly electricity consumption has three seasonalities: daily (24h), weekly (168h), yearly (8760h). SARIMA handles only one. Prophet can handle multiple but scales poorly to thousands of series.

Missing values and anomalies. A missing sensor reading is information (the sensor turned off), not NaN. Linear interpolation destroys this signal. Proper handling depends on the missingness mechanism.

Cold start. A new SKU in a 50,000‑item assortment has no history, yet a forecast is needed. Standard approaches fail; cross‑learning or feature‑based methods are required.

Why is model selection critical for your data?

Prophet (Meta) – a solid start for business data with clear seasonality and holidays. Fast setup, interpretable, built‑in outlier detection. Fails on irregular patterns and does not scale beyond ~10k series without parallelization.

Gradient boosting on features (LightGBM, XGBoost) – often underestimated. Engineer lags (t‑1, t‑7, t‑28), rolling means, day‑of‑week, holidays. The model trains on all series simultaneously, solving cold start via transfer learning. MAPE in retail often beats neural nets with proper feature engineering.

TFT (Temporal Fusion Transformer) – a transformer designed for interpretable forecasting with covariates. Built‑in variable selection, temporal attention, quantile outputs. Available in pytorch‑forecasting. Requires ~10,000+ records per series for stable training.

PatchTST – splits the series into patches (like ViT for images), capturing local patterns better than classic transformers. Excellent for long‑horizon forecasting (96–720 steps ahead).

N‑HiTS, N‑BEATS – attention‑free neural architectures, faster than TFT, competitive accuracy. N‑BEATS won the M4/M5 benchmarks for tasks without covariates.

How do we deploy TFT in production?

A typical pipeline via pytorch‑forecasting:

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="sales",
    group_ids=["store", "sku"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,  # 120 days
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,  # 28 days
    static_categoricals=["store_type", "category"],
    time_varying_known_reals=["price", "promo_flag"],
    time_varying_unknown_reals=["sales"],
    target_normalizer=GroupNormalizer(groups=["store", "sku"], transformation="softplus"),
)

A common mistake: the default target_normalizer (StandardScaler) breaks predictions for series with zero values (no sales on weekends). GroupNormalizer with transformation="softplus" is the correct choice for count data.

Case study: retail demand forecasting

A chain of 120 stores, 8,000 SKUs, 28‑day forecast horizon. The original system: SARIMA per series, MAPE 18.4%, retraining cycle – 6 hours. We replaced it with TFT on PyTorch + pytorch‑forecasting: a single model for all series, MAPE 11.2%, retraining – 40 minutes on an A10G. Feature importance via variable selection revealed that day_before_holiday influences more than the holiday date itself. Annual savings on inference alone exceeded $50,000.

Step‑by‑step configuration

1. Data collection and preparation. Handle missing values (mark NaN, interpolate only for technical failures), aggregate to required frequency, engineer covariates (holidays, promotions, prices).
2. Create TimeSeriesDataSet. Set group_ids (store + SKU), time index, forecast horizon. Choose target_normalizer based on target distribution.
3. Train a baseline. Prophet or LightGBM first – to understand complexity.
4. Train TFT. Use TemporalFusionTransformer with loss=QuantileLoss(), tune learning rate and hidden layer sizes.
5. Validate and interpret. Walk‑forward test, analyze variable selection, build attention heatmaps.

How to properly evaluate forecast quality?

RMSE alone is misleading – it over‑penalizes large values. Our standard set:

• MAPE – interpretable, unstable near zero.
• sMAPE – symmetric, avoids division by small numbers.
• MASE (Mean Absolute Scaled Error) – normalized relative to a naive seasonal forecast, ideal for comparing series of different scales.
• Pinball loss – for probabilistic forecasting, inventory management.

We guarantee a model card with these metrics on the validation set and walk‑forward results on at least 6 months of history.

What deliverables do you receive?

• Documentation of chosen architecture and hyperparameter rationale.
• Reproducible training and inference pipeline (Docker + CI/CD + Airflow/Prefect).
• Committed code with unit tests for key components.
• Team training: retraining, output interpretation, deployment of new versions.
• 3 months of post‑delivery support (consultations, bug fixes, fine‑tuning).

The model is deployed via FastAPI or Triton Inference Server. Retraining is scheduled (e.g., weekly) via Airflow with drift validation and automatic rollback if metrics deteriorate.

Process and timeline

We start with EDA: visualization, ADF test, STL decomposition, analysis of missing values and outliers. This takes 2–3 days but often reveals systemic data issues that block forecasting. Then we build a baseline (naive seasonal, Prophet), engineer features for LightGBM, and select a neural architecture if needed. Walk‑forward validation with a realistic horizon. Deployment via API with automatic retraining scheduled via Airflow or Prefect.

Timeline: MVP forecast on one data type – 3–6 weeks. Hierarchical forecasting system with automation – 2–5 months. Cost is calculated individually based on data volume, number of series, and required accuracy.

Our team consists of certified ML engineers (AWS ML Specialty, GCP Professional ML Engineer) with 5+ years on the market and over 50 completed forecasting projects. Contact us for a free analysis of your data – we will assess the task and provide initial recommendations within 1–2 days. Request a consultation to ensure your forecasts work in production, not just in a notebook.