Page 1 – Financial Time-Series: Structure, Features & Challenges

1.1 Nature of Financial Time-Series

Financial time-series differ from standard ML datasets. They are noisy, non-stationary, heavy-tailed, and exhibit temporal dependencies that evolve across market regimes.

1.2 Key Stylised Facts

• Volatility clustering: high-volatility periods follow high-volatility periods.
• Fat tails: extreme returns occur more than Gaussian models predict.
• Mean reversion: some assets show pullback towards long-term equilibrium.
• Leverage effects: volatility increases when prices fall.
• Regime shifts: structural breaks (policy changes, crises).

1.3 Types of Financial Forecasting Problems

• Return forecasting – directional prediction or regression.
• Volatility forecasting – risk estimation for derivatives and VaR.
• Liquidity forecasting – bid–ask spreads, depth, order arrivals.
• Price level forecasting – FX, commodities, rates.

1.4 Signal-to-Noise Ratio (SNR)

Financial data has extremely low SNR. Most short-term returns are noise. AI must identify small, stable patterns that survive costs and slippage.

1.5 Data Granularity

• Daily / Monthly: macro, broad trends.
• Intraday (1h, 5m): tactical strategies.
• Tick / Microsecond: HFT and microstructure modelling.

1.6 Challenges Unique to Time-Series ML

• Non-stationarity and changing distributions.
• Data leakage over time.
• Regime dependency of model behaviour.
• Computational intensity for intraday horizons.

Page 2 – Forecasting Models: Statistical & Machine Learning Approaches

2.1 Classical Statistical Models

Widely used baseline models include:

• ARIMA (AutoRegressive Integrated Moving Average)
• ARMA, AR, MA
• Exponential smoothing and Holt–Winters
• GARCH-family volatility models

2.2 Volatility Models

• GARCH – time-varying volatility.
• EGARCH – asymmetries / leverage effects.
• GJR-GARCH – threshold effects.
• Stochastic volatility – latent state-space models.

2.3 ML Regression Models

Machine learning models for forecasting include:

• Random Forest Regression
• Gradient Boosting (XGBoost, LightGBM)
• Elastic Net / LASSO for sparse predictors
• SVM regression

These models perform well for engineered features (lags, technical indicators, macro signals).

2.4 Deep Learning Architectures

• LSTMs – capture long-term dependencies; widely used in FX and crypto forecasting.
• GRUs – computationally efficient sequential models.
• Temporal CNNs (TCNs) – dilated convolutions for long-range memory.
• Transformers for Time-Series – attention over temporal windows (Temporal Fusion Transformer, Informer, FEDformer).

2.5 Feature Engineering Techniques

• Lagged returns, rolling windows, moving averages.
• Volatility and range-based indicators (ATR, Parkinson volatility).
• Microstructure features (order-book imbalance, spread).
• Calendar and event features (options expiry, announcements).

2.6 Forecasting Horizon Considerations

• Short horizon: predictive signals are tiny and fragile.
• Medium horizon: stronger macro/technical structure.
• Long horizon: regime-driven, less noise, but harder to label.

Page 3 – Machine Learning for Algorithmic Trading

3.1 Trading Strategy Types

• Mean-reversion – price deviations revert to fundamental value.
• Momentum / trend following – returns persist in direction.
• Statistical arbitrage – pairs trading, cointegration.
• Cross-sectional signals – ranking assets based on features.
• Market making – capturing spreads via quoting and inventory control.

3.2 Signal Engineering

Machine learning signals for trading include:

• Price-based signals: lags, volatility, autocorrelations.
• Order-book features: queue sizes, imbalance, depth.
• Event-driven signals: macro news, earnings announcements.
• NLP signals: sentiment from financial news & social platforms.

3.3 Execution Algorithms

Execution models focus on minimizing market impact:

• VWAP – trade proportional to market volume.
• TWAP – evenly paced execution.
• POV – trades based on participation rate.
• Optimal execution – Almgren–Chriss model for cost minimisation.

3.4 Reinforcement Learning for Trading

RL frames trading as a sequential decision problem:

• State: prices, positions, market conditions.
• Action: buy, sell, hold, adjust size.
• Reward: P&L adjusted for risk and costs.

Constraints:

• High risk of overfitting to backtests.
• High sensitivity to regime changes.
• Regulatory scrutiny for autonomy in financial markets.

3.5 Backtesting & Simulation

Trading models require robust backtesting:

• Train/validation/test split by time.
• Realistic transaction cost modelling.
• Bootstrap and Monte Carlo scenario testing.
• Avoiding look-ahead and survivorship biases.

Page 4 – Market Microstructure & High-Frequency AI Models

4.1 Market Microstructure

Microstructure studies how trading mechanisms, order books, and liquidity shape price formation. AI is increasingly used to model:

• Order arrival intensities
• Bid–ask spread dynamics
• Order-book imbalance
• Price impact of trades

4.2 Order-Book Features

• Top-of-book quotes (best bid, best ask)
• Depth at each level
• Queue position and order lifetime
• Volume imbalance indicators

4.3 High-Frequency Forecasting Models

Popular models include:

• Hawkes processes – self-exciting point processes for order flow.
• DeepLOB – deep learning on limit order books.
• Temporal GNNs – modelling order-book graphs.
• TCN / LSTM hybrids – short-term directional prediction.

4.4 Price Impact Modelling

AI models estimate:

• Instantaneous impact of trades.
• Permanent vs temporary impact.
• Nonlinear reaction to volume and market state.

4.5 Latency & Execution Constraints

• Decision latency (microseconds for HFT).
• Co-location requirements.
• Hardware optimisation.
• Real-time risk checks.

Page 5 – Limitations, Risks & Summary

5.1 Limitations of Time-Series AI in Finance

• Low predictability of short-term returns.
• Extreme sensitivity to regime changes.
• High model variance due to noise.
• Difficulty generalising to unseen market states.

5.2 Backtest Overfitting & P-Hacking

With many features and models, it’s easy to “discover” spurious patterns that never work in real markets.

• Need strict out-of-sample validation.
• Use statistical tests such as White’s Reality Check.
• Apply cross-validation with temporal blocking.

5.3 Real-World Constraints

• Transaction costs & slippage.
• Liquidity constraints.
• Market impact.
• Execution risk.
• Regulatory reporting obligations.

5.4 Ethical & Regulatory Considerations

• Models must not destabilise markets.
• Algorithmic trading requires audit trails and kill-switches.
• Risk committees must approve model behaviour.

5.5 Summary

Module 3 provided a comprehensive overview of AI techniques for financial time-series forecasting and algorithmic trading, including:

• Key stylised facts and forecasting challenges.
• Statistical, ML, and deep learning forecasting approaches.
• Trading signals, execution algorithms, and RL frameworks.
• Market microstructure modelling and high-frequency techniques.
• Limitations, risks, and regulatory considerations.

Module 4 will cover NLP, alternative data, and unstructured signals in financial AI.