AI System for Optimal Position Sizing

The Problem: Traders Lose Capital Due to Incorrect Sizing

Even a profitable trading strategy breaks down with the wrong position size. Too aggressive a size — and one bad trade sends the portfolio into a deep drawdown. Too conservative — returns are below potential. We develop AI systems to calculate and adapt position size in real time, solving this dilemma. Our experience: 5+ years in algorithmic trading and over 30 projects in trading system development. Average savings from preventing one major drawdown in our projects — $20,000 in the first year. Another example: one client avoided a $50,000 loss due to timely position size adjustment.

Problems We Solve

Fixed Risk Percentage Ignores Signal Quality

If a strategy yields 55% accurate predictions on some data and 80% on others, position size should reflect that difference. Without adaptation, you either underperform on good trades or risk too much on bad ones.

Kelly Criterion Requires Accurate Estimates

Full Kelly (0.325 of capital at win rate 55% and R=2) will ruin you with the slightest estimation error. Half-Kelly (0.5 of Kelly) is a common compromise, but it still ignores volatility and correlations. Our RL agents learn to find optimal size without manual coefficient tuning.

Volatility and Drawdown Are Not Accounted For

The same strategy in a calm market versus high volatility requires different risk levels. When the portfolio is in a drawdown, position size must decrease to preserve capital. We implement volatility-adjusted sizing and drawdown-adjusted sizing (anti-Martingale) with an automatic circuit breaker when thresholds are exceeded.

How Our AI System Adapts Position Size

The system takes as input an array of features: signal confidence (probability score), 10-day realized volatility, current drawdown from peak, and macro regime (expansion/contraction). Based on these, an ML model or RL agent outputs the optimal position size as a percentage of capital. This happens on each tick or on a schedule with latency p99 <5 ms.

Why an RL Agent is More Effective Than Kelly Criterion

The Kelly Criterion is theoretically optimal for long-term growth, but in practice it leads to 50% drawdown due to inaccurate probability estimates. An RL agent, trained on historical data, automatically chooses conservative sizing in uncertain situations and aggressive sizing when confidence is high. In our projects, the RL approach improves Sharpe by 1.3x compared to half-Kelly. At the same time, the risk of critical drawdown is reduced by 20%. Internal research

How We Do It: Tech Stack and a Case Study

For one task, we developed an RL agent in PyTorch with state: signal confidence, 10-day realized volatility, current drawdown, macro regime. Discrete action space: 0%, 0.5%, 1%, 1.5%, 2%, 3% risk per trade. Reward: PnL with a penalty for exceeding drawdown >20%. Results: Sharpe ratio increased from 1.2 (fixed fractional 1%) to 1.6, maximum drawdown decreased from 35% to 22%.

Tech stack: PyTorch for model, Ray RLlib for training, pgvector for storing market state embeddings, Triton Inference Server for inference with latency p99 <5 ms. Model quantized (INT8) for low latency.

Comparison of Sizing Methods

The RL agent delivers the best metrics but requires more data and computational resources. Backtesting results available on request.

Process and Scope of Work

1. Analytics: collect historical data, identify patterns and correlations. Define target metrics (Sharpe, max drawdown, recovery factor).
2. Design: select sizing architecture (volatility-adjusted, RL, risk parity). Design state/action/reward.
3. Implementation: develop ML pipeline (feature engineering, training, validation). Integrate with broker API or trading terminal.
4. Testing: backtesting on out-of-sample data, Monte Carlo simulation (10,000+ trajectories). Stress-testing on crisis periods.
5. Deployment and monitoring: deploy model on server, set up monitoring of metrics (volatility, drawdown, performance attribution).

Timelines and Deliverables

Each delivery includes: Python source code, model in ONNX format, documentation of method and parameters, consultations during integration. Cost is calculated individually — we will evaluate your project based on your terms of reference.

Common Mistakes in Implementation

• Using full Kelly without volatility adjustment — a quick way to blow up.
• Ignoring transaction costs and slippage — they can eat profits, especially with frequent position recalculations.
• Overfitting the RL agent to historical data — always test out-of-sample on different market regimes.

Get a Consultation

Contact us to discuss your strategy and design an optimal sizing system. We guarantee a personalized approach and transparency at all stages. Get a consultation with an engineer: we'll show you how adaptive sizing can improve your metrics.

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.