Custom AI Model for Asset Price Prediction

Why Most AI Price Forecasting Models Fail?

Traders and funds often invest resources in developing complex models, but on real markets they show losses. The reason lies in three traps: look-ahead bias, ignoring transaction costs, and overfitting to historical data. For example, a simple moving average model may show a Sharpe ratio of 1.5 in backtest, but in live trading without accounting for slippage and commissions, this drops to 0.3. We have encountered projects where a model with IC 0.08 on validation showed Sharpe 0.2 in production — due to ignoring slippage. Our approach eliminates such surprises. We solve these issues through purged walk-forward validation, a realistic transaction cost model (Almgren-Chriss), and strict factor selection control.

How to Choose the Forecasting Horizon?

The practical goal is not the exact price in N days, but a signal with positive expected value after transaction costs. Even a model with MAPE 3% on S&P500 stocks is useless if the strategy's Sharpe ratio is < 0. The horizon determines the signal type:

• Intraday (minutes-hours): microstructure signals, order flow imbalance — typical return 0.5–1.5% per trade.
• Short-term (1-5 days): momentum, mean reversion — average IC 0.05–0.08.
• Medium-term (1-4 weeks): earnings, macro catalysts — IC can reach 0.12.
• Long-term (months): fundamental valuation, factor exposure — more stable but requires higher accuracy.

The optimal horizon depends on the instrument's liquidity and rebalancing frequency. For less liquid assets, shorter horizons are less reliable.

What is Purged Walk-Forward Validation?

Correct validation is key to a realistic backtest. We use purged walk-forward cross-validation:

• Training: t=0 to t=T
• Purge gap: T to T+embargo (eliminates look-ahead from overlapping labels)
• Test: T+embargo to T+embargo+H
• Embargo period: usually equal to the forecast horizon

Embargo period ensures that future information does not leak into the training set. This is critical for time series. Metrics: IC (Information Coefficient) — correlation between predicted and actual return ranks. IC > 0.05 is weak, IC > 0.10 is good. ICIR (IC Information Ratio) — signal stability. Strategy Sharpe ratio from the signal is the main practical metric. Efficient Market Hypothesis states markets are efficient, but in practice micro-anomalies exist and can be identified with correct validation.

For model selection, consider the volume and structure of data: if many instruments — LightGBM ranking, if a single time series — LSTM, if multi-instruments with known events — Temporal Fusion Transformer. With limited data, start with LightGBM.

Model Features and Architecture

Price-based (technical analysis):

• Returns: log returns for 1, 5, 10, 21 trading days.
• Momentum: 12-1 month momentum (Jegadeesh-Titman factor).
• RSI, MACD, Bollinger Band width — oscillators as functions of price.
• Volatility: realized volatility for 5/21/63 days.

Volume-based:

• Volume relative to 20-day average.
• Price × Volume (dollar volume).
• On-Balance Volume (OBV).
• VWAP deviation.

Fundamental (for stocks):

• P/E, P/B, EV/EBITDA.
• EPS growth YoY.
• Revenue growth.
• Debt/Equity.

Alternative data:

• Sentiment from Twitter/Reddit (NLP score).
• Google Trends for consumer stocks.
• Satellite imagery (retail parking lots, commodity stores).
• Job postings growth (Glassdoor, LinkedIn).

Comparison of main modeling approaches:

LightGBM trains 10x faster than LSTM on tabular data — an advantage for rapid prototyping. For ranking tasks, we use LGBMRanker with objective='lambdarank'. Example configuration:

import lightgbm as lgb

model = lgb.LGBMRanker(
    objective='lambdarank',
    n_estimators=500,
    learning_rate=0.05,
    max_depth=6
)

For single instrument time series, we use LSTM with 60 days of history:

model = Sequential([
    LSTM(64, return_sequences=True, input_shape=(60, n_features)),
    Dropout(0.2),
    LSTM(32),
    Dropout(0.2),
    Dense(1)
])

Temporal Fusion Transformer — the best choice when known future covariates (earnings dates, macro events) and 100+ instruments are available.

Model quality is assessed not only by IC. We use a comprehensive set of metrics:

From Model to Trading Strategy

Model → signal → position → PnL — a chain with multiple stages of loss:

1. Signal generation: ranking score across stock universe (typically 500-1000 instruments).
2. Portfolio construction: mean-variance optimization (Markowitz) or equal-weight deciles. Typical number of positions 20-50.
3. Risk management: limits on sector/factor exposure, max position size 5%.
4. Transaction cost model: bid-ask spread + market impact (Almgren-Chriss) — accounts for slippage, often 10-30 bps.
5. Backtesting: with real TC and slippage — key! We use Zipline / Backtrader or custom backtester.

Common mistakes: survivorship bias (training only on existing stocks), look-ahead bias in fundamental data (use point-in-time), ignoring transaction costs. We document every assumption.

What's Included

• Documentation: dashboard with metrics (IC, Sharpe), model description, reproducible code.
• Model access: REST API or Python package with documentation.
• Team training: workshop on operation and retraining.
• Support: 3 months after deployment, including drift monitoring.

Get an engineer consultation for your project — we'll assess data and timelines for free.

Our Results

We have built models for several hedge funds and prop trading teams. Average savings on transaction costs are 20-30% compared to naive benchmarks. We guarantee IC > 0.05 on out-of-sample, Sharpe ratio > 1.0 after TC. Certified in AWS and GCP ML. Experience with LightGBM and PyTorch. Contact us for a project evaluation — we'll calculate timelines and cost for free.

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.