Developing a Financial AI Model with Transformers

We often get requests from trading funds and fintech companies: LSTM models fail to capture long-term dependencies, while Prophet ignores exogenous events. Transformer architectures, which revolutionized sequence processing, solve these problems with the self-attention mechanism. It allows the model to explicitly focus on relevant points in history—for instance, a historical financial crisis can influence today’s forecast despite a gap of a thousand time steps.

Problems We Solve

Financial time series have unique characteristics: non-stationarity, heteroscedasticity, multi-period seasonality, and regime changes. Standard RNN/CNN approaches require manual feature engineering and do not scale to multiple instruments. Another challenge is multimodality: price, news, macroeconomic indicators, options data. Transformers naturally integrate heterogeneous data types through cross-attention.

How We Do It: The Temporal Fusion Transformer Case

One effective approach is the Temporal Fusion Transformer (TFT). We implemented it for a hedge fund: 100+ instruments, daily data, 5-day forecast horizon. TFT includes a Variable Selection Network that automatically picks relevant features, and quantile forecasting (p10/p50/p90) for uncertainty estimation. On the test period, TFT outperformed vanilla Transformer by 18% MAPE and 12% Sharpe ratio on a simulated strategy. For reproduction we use the pytorch-forecasting library, which provides ready-made implementations.

from pytorch_forecasting import TemporalFusionTransformer, TimeSeriesDataSet

training = TimeSeriesDataSet(
    data,
    time_idx="time_idx",
    target="return",
    group_ids=["ticker"],
    max_encoder_length=60,
    max_prediction_length=5,
    time_varying_known_reals=["vix", "dollar_index", "yield_10y"],
    time_varying_unknown_reals=["return", "volume", "rsi", "atr"],
)
tft = TemporalFusionTransformer.from_dataset(training)

Can You Use a Vanilla Transformer for Financial Series?

Yes, but with caveats. A vanilla Transformer with causal masking works for single-asset forecasting with a context length of up to 100 steps. However, it does not handle multiple instruments and lacks interpretability. For real projects we recommend specialized architectures.

What is the Temporal Fusion Transformer and Why is It Better?

TFT is a hybrid architecture: GRU for local patterns + self-attention for long-term dependencies + Variable Selection Network for feature selection. It outputs quantile forecasts, which is critical for risk management. On benchmarks, TFT consistently outperforms vanilla Transformer by 15-20% on multivariate financial series.

Our Work Process

We implement the full cycle:

What Is Included in Turnkey Development

The standard package includes: data preparation (cleaning, aggregation, feature engineering), architecture selection and customization, training with automatic hyperparameter tuning (Optuna, Weights & Biases), integration with your infrastructure, API documentation, and team training. Optionally, we can set up data drift monitoring and automatic retraining.

Timelines and Cost

Development timelines: from 3 weeks for a single-asset baseline to 3-5 months for a custom multi-asset solution with news fusion. Cost is calculated individually—depends on the number of instruments, data type, and latency requirements. The payback period for such a model averages 8-12 months. Contact us for a preliminary assessment of your project.

Typical Mistakes When Training Financial Transformers

• Using non-stationary series without differencing or Box-Cox.
• Missing causal masking—leaking future information.
• Overfitting on one-year data without accounting for regime changes.
• Ignoring outliers—a spike in VIX can distort attention.
• Using too long a context (>250 steps) without sparse attention.

We ensure development quality: certified engineers with 10+ years of experience in financial ML have delivered 50+ Transformer-based projects. To discuss your task, write to us—we’ll help select the architecture and estimate the project. Savings on computational resources through proper quantization reach 30%.

Vaswani et al., "Attention is All You Need" (original article)

# Warmup + cosine decay
def lr_lambda(step):
    if step < warmup_steps:
        return step / warmup_steps
    progress = (step - warmup_steps) / (total_steps - warmup_steps)
    return 0.5 * (1 + math.cos(math.pi * progress))

Get a consultation from our engineer to discuss your task and possible architecture.

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.