LSTM for Financial Time Series: Architecture and Validation

LSTM for Financial Time Series: Architecture and Validation

Imagine this: you trained an LSTM on five years of daily data, got 68% accuracy on the test set. In production, the model shows 49% — worse than random. Typical mistake: data leakage during normalization or incorrect validation. We deploy production-ready LSTM architecture for financial time series based on real projects with multi-asset portfolios and walk-forward validation. Our team has 10+ years of experience in AI/ML for finance, implementing 30+ models for hedge funds and brokers. We guarantee no lookahead bias and experiment reproducibility. We use PyTorch and Hugging Face Transformers, train on an A100 GPU cluster, monitor via MLflow and Weights & Biases. Hyperparameter optimization is done with Optuna, validation is strict walk-forward with an embargo period to eliminate leakage. Result: stable Information Coefficient (IC) > 0.05 and ICIR > 1.5 on out-of-sample test. Development cost depends on model complexity and data volume — final price is discussed after analysis. Estimated duration for a single-asset solution is 2 to 3 weeks of team work, multi-asset with attention is 8 to 10 weeks.

Why LSTM, not gradient boosting?

LSTM wins when sequence of events is more important than aggregates, and nonlinear time patterns are explicit. LightGBM with lag features often beats LSTM on small datasets (<10,000 observations). But on multivariate series (multiple instruments simultaneously) and complex cross-asset dependencies, LSTM offers an advantage. The architecture was first described in the paper Long Short-Term Memory (Hochreiter & Schmidhuber). LSTM is the base architecture.

Model Architecture

import torch
import torch.nn as nn

class FinancialLSTM(nn.Module):
    def __init__(self, input_size, hidden_size=128, num_layers=2, dropout=0.2):
        super().__init__()
        self.lstm = nn.LSTM(
            input_size=input_size,
            hidden_size=hidden_size,
            num_layers=num_layers,
            batch_first=True,
            dropout=dropout
        )
        self.attention = nn.MultiheadAttention(hidden_size, num_heads=8)
        self.fc = nn.Linear(hidden_size, 1)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        lstm_out, _ = self.lstm(x)  # [batch, seq_len, hidden]
        # Self-attention over time dimension
        attn_out, _ = self.attention(lstm_out, lstm_out, lstm_out)
        # Last step or attention-weighted pool
        out = self.fc(self.dropout(attn_out[:, -1, :]))
        return out

Input data (seq_len × n_features): OHLCV, normalized by rolling window, technical indicators (RSI, MACD, ATR, Bollinger). For multi-asset — concatenation along feature dimension. Implementation is available in PyTorch LSTM.

Preprocessing and Normalization

Critically important: normalization without lookahead bias. We use rolling window normalization:

def rolling_normalize(X, window=252):
    mu = X.rolling(window).mean()
    sigma = X.rolling(window).std()
    return (X - mu) / (sigma + 1e-8)

Price returns instead of prices: raw prices are non-stationary, log returns are stationary:

returns = np.log(prices / prices.shift(1)).dropna()

Sequence generation:

def create_sequences(data, seq_len=60, horizon=5):
    X, y = [], []
    for i in range(len(data) - seq_len - horizon):
        X.append(data[i:i+seq_len])
        y.append(data[i+seq_len+horizon-1, 0])
    return np.array(X), np.array(y)

Training and Regularization

How to tune hyperparameters for financial LSTMs?

Sequence length: 20–60 days for daily data, 50–200 for hourly. Hidden size: 64–256. Layers: 2–3 (deeper is usually worse on financial data). Dropout: 0.1–0.4. Batch size: 32–128. Regularization: temporal dropout, feature noise, L2 weight decay (1e-4 to 1e-3). Optimizer: AdamW with cosine annealing LR scheduler. Early stopping on validation loss on a 20% holdout.

For a portfolio of N instruments, we use Cross-sectional LSTM with parallel processing of all instruments and cross-attention between them to capture correlation patterns (oil → oil stocks, DXY → EM assets).

Validation Without Data Leakage

Walk-forward with embargo:

embargo_size = horizon
train_end = int(0.6 * len(data))
embargo_end = train_end + embargo_size
val_end = int(0.8 * len(data))

Metrics: Directional Accuracy, Information Coefficient (spearman correlation), ICIR (IC / std(IC) — stability; ICIR > 1.5 is considered good).

Comparison of Normalization Methods

LSTM vs Transformer for Finance

For short sequences (< 100 steps), LSTM often matches Transformer with significantly less data requirements.

What Is Included in the Work

• Baseline single-asset model with built pipeline and documentation
• Multi-asset architecture with cross-attention and walk-forward validation
• Hyperparameter optimization (Optuna) with logs in MLflow
• Docker deployment with Triton Inference Server and monitoring in Prometheus
• Training of the operations team and handover of the model card

Each stage is accompanied by reports and code comments. We don't just deliver weights — we hand over a reproducible experiment.

Process and Timelines

• Analysis — data collection and visualization, defining the forecast horizon.
• Design — selection of architecture (LSTM/Transformer, single/multi-asset).
• Implementation — writing pipeline, training baseline, optimization.
• Testing — walk-forward validation, stress testing on anomalies.
• Deployment — packaging in Docker, deploying on a GPU server, monitoring.

Timelines: single-asset baseline — 2 to 3 weeks; multi-asset model with attention and production pipeline — 8 to 10 weeks. Cost is calculated individually.

Request a consultation for a preliminary assessment of your dataset — we will analyze it in 1–2 days. Contact us to discuss the model architecture and timelines.

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.