Generative AI for Trading and Asset Management is an essential guide to understand how generative AI has emerged as a transformative force in the realm of asset management, particularly in the context of trading, due to its ability to analyze vast datasets, identify intricate patterns, and suggest complex trading strategies. Practically, this book explains how to utilize various types of AI: unsupervised learning, supervised learning, reinforcement learning, and large language models to suggest new trading strategies, manage risks, optimize trading strategies and portfolios, and generally improve the productivity of algorithmic and discretionary traders alike. These techniques converge into an algorithm to trade on the Federal Reserve chair's press conferences in real time.
Written by Hamlet Medina, chief data scientist Criteo, and Ernie Chan, founder of QTS Capital Management and Predictnow.ai, this book explores topics including: - How large language models and other machine learning techniques can improve productivity of algorithmic and discretionary traders from ideation, signal generations, backtesting, risk management, to portfolio optimization - The pros and cons of tree-based models vs neural networks as they relate to financial applications. How regularization techniques can enhance out of sample performance - Comprehensive exploration of the main families of explicit and implicit generative models for modeling high-dimensional data, including their advantages and limitations in model representation and training, sampling quality and speed, and representation learning. - Techniques for combining and utilizing generative models to address data scarcity and enhance data augmentation for training ML models in financial applications like market simulations, sentiment analysis, risk management, and more. - Application of generative AI models for processing fundamental data to develop trading signals. - Exploration of efficient methods for deploying large models into production, highlighting techniques and strategies to enhance inference efficiency, such as model pruning, quantization, and knowledge distillation. - Using existing LLMs to translate Federal Reserve Chair's speeches to text and generate trading signals.
Generative AI for Trading and Asset Management earns a well-deserved spot on the bookshelves of all asset managers seeking to harness the ever-changing landscape of AI technologies to navigate financial markets.
Table of Contents
Table of contents
PREFACE
ACKNOWLEDGMENTS
About the Authors
PART I: GENERATIVE AI FOR TRADING AND ASSET MANAGEMENT: A LOW-CODE INTRODUCTION
1. NO-CODE GENERATIVE AI FOR BASIC QUANTITATIVE FINANCE
1.1 RETRIEVING HISTORICAL MARKET DATA
1.2 COMPUTING SHARPE RATIO
1.3 DATA FORMATTING AND ANALYSIS
1.4 TRANSLATING MATLAB CODES TO PYTHON CODES
1.5 CONCLUSION
2. NO-CODE GENERATIVE AI FOR TRADING STRATEGIES DEVELOPMENT
2.1 CREATING CODES FROM A STRATEGY SPECIFICATION
2.2 SUMMARIZING A TRADING STRATEGY PAPER AND CREATING BACKTEST CODES FROM IT
2.3 SEARCHING FOR A PORTFOLIO OPTIMIZATION ALGORITHM BASED ON MACHINE LEARNING.
2.4 EXPLORE OPTIONS TERM STRUCTURE ARBITRAGE STRATEGIES
2.5 CONCLUSION
2.6 EXERCISES
3. WHIRLWIND TOUR OF ML IN ASSET MANAGEMENT
3.1 UNSUPERVISED LEARNING
3.1.1 HIERARCHICAL RISK PARITY (HRP)
3.1.2 PRINCIPAL COMPONENT ANALYSIS (PCA)
3.1.3 CLUSTER-BASED FEATURE SELECTION (CMDA)
3.1.4 HIDDEN MARKOV MODEL (HMM)
3.2 SUPERVISED LEARNING
3.2.1 LINEAR AND LOGISTIC REGRESSIONS
3.2.2 L1 AND L2 REGULARIZATIONS
3.2.3 HYPERPARAMETER OPTIMIZATION, VALIDATION AND CROSS-VALIDATION
3.2.4 PERFORMANCE METRICS
3.2.5 CLASSIFICATION AND REGRESSION TREES, RANDOM FOREST, AND BOOSTED TREES
3.2.6 NEURAL NETWORKS
3.2.7 RECURRENT NEURAL NETWORK
3.3 DEEP REINFORCEMENT LEARNING
3.4 DATA ENGINEERING
3.4.1 UNIQUE COMPANY IDENTIFIERS
3.4.2 DIVIDEND AND SPLIT ADJUSTMENTS
3.4.3 SURVIVORSHIP BIAS
3.4.4 LOOK-AHEAD BIAS
3.5 FEATURES ENGINEERING
3.5.1 STATIONARITY
3.5.2 MERGING TIME SERIES WITH DIFFERENT FREQUENCIES
3.5.3 TIME-SERIES VS CROSS-SECTIONAL FEATURES
3.5.4 VALIDATING THIRD-PARTY FEATURES
3.5.5 GENERATIVE AI AS A FEATURE GENERATOR
3.5.6 FEATURES IMPORTANCE RANKING AND SELECTION
3.6 CONCLUSION
PART II: DEEP GENERATIVE MODELS FOR TRADING AND ASSET MANAGEMENT
4. UNDERSTANDING GENERATIVE AI
4.1 WHY GENERATIVE MODELS
4.2 DIFFERENCE WITH DISCRIMINATIVE MODELS
4.3 HOW CAN WE USE THEM?
4.3.1 PROBABILITY DENSITY ESTIMATION
4.3.2 GENERATING NEW DATA
4.3.3 LEARNING NEW DATA REPRESENTATIONS.
4.4 TAXONOMY OF GENERATIVE MODELS
4.5 CONCLUSION
5. DEEP AUTO-REGRESSIVE MODELS FOR SEQUENCE MODELING
5.1 REPRESENTATION COMPLEXITY
5.2 REPRESENTATION AND COMPLEXITY REDUCTION
5.3 A SHORT TOUR OF KEY MODEL FAMILIES
5.3.1 LOGISTIC REGRESSION MODEL
5.3.1.1 Sampling from FVSN
5.3.2 MASKED AUTO ENCODER FOR DENSITY ESTIMATION (MADE)
5.3.3 CAUSAL MASKED NEURAL NETWORK MODELS
5.3.3.1 WaveNet
5.3.4 RECURRENT NEURAL NETWORKS (RNN)
Practical Considerations and Limitations
5.3.5 TRANSFORMERS
5.3.5.1 Attention mechanism
5.3.5.2 Scaled Dot-Product Attention
5.3.5.3 From Self-Attention To Transformers
5.3.5.4 Positional Encodings
5.3.5.5 MultiHeaded Attention
5.3.5.6 The Feed-Forward Layer
5.3.5.7 Add & Norm blocks
5.3.5.8 The Transformer Encoder layer
5.3.5.9 The Complete Transformer Encoder
5.3.5.10 Model Objective.
5.3.6 FROM NLP TRANSFORMER TO THE TIME SERIES TRANSFORMERS
5.3.6.1 Discretizing Time Series Data: The Chronos Approach.
5.3.6.2 Continuous Input for Transformers: The Lag-Llama Approach
5.3.6.2.1 Innovations and approaches
5.4 MODEL FITTING
5.5 CONCLUSIONS
6. DEEP LATENT VARIABLE MODELS
6.1 INTRODUCTION
6.2 LATENT VARIABLE MODELS
6.3 EXAMPLES OF TRADITIONAL LATENT VARIABLE MODELS
6.3.1 FACTOR ANALYSIS
6.3.2 PROBABILISTIC PRINCIPAL COMPONENT ANALYSIS
Example: Comparing PCA and Factor Analysis for Latent Space Recovery
Advantages of Probabilistic Approaches (PPCA/FA) over PCA
6.3.3 GAUSSIANS MIXTURE MODELS
6.3.3.1 Gaussian Mixture Model (GMM) for Market Regime Detection
Example: Low and High Volatility Regimes
6.3.4 DEEP LATENT VARIABLE MODELS
6.4 LEARNING
6.4.1 TRAINING OBJECTIVE
6.4.2 THE VARIATIONAL INFERENCE APPROXIMATION
How to Choose the Proposal Distribution
Amortized inference.
6.4.3 OPTIMIZATION
6.4.3.1 The Likelihood gradient, REINFORNCE
6.4.3.2 Reparameterization trick
6.4.4 MIND THE GAP!
6.5 VARIATIONAL AUTO ENCODERS (VAE)
6.6 VAES FOR SEQUENTIAL DATA AND TIME SERIES
6.6.1 EXTENDING VAES FOR TIME SERIES
Sequential Encoders and Decoders
Superposition of Time Serie Components
6.6.1.1 TimeVAE: A flexible VAE for Time Series Generation
Architecture of TimeVAE
6.7 CONCLUSION
7. FLOWS MODELS
7.1 INTRODUCTION
7.2 TRAINING
7.3 LINEAR FLOWS
7.4 DESIGNING NON LINEAR FLOWS
7.5 COUPLING FLOWS
7.5.1 NICE: NONLINEAR INDEPENDENT COMPONENTS ESTIMATION
7.5.2 REAL-NVP: NON VOLUME PRESERVING TRANSFORMATION
7.6 AUTOREGRESSIVE FLOWS
7.7 CONTINUOUS NORMALIZING FLOWS
7.8 MODELING FINANCIAL TIME SERIES WITH FLOW MODELS.
7.8.1 TRANSITIONING FROM IMAGE DATA TO TIME SERIES DYNAMICS
7.8.2 ADAPTING FLOWS FOR TIME SERIES.
7.8.3 CASE STUDY: A PRACTICAL EXAMPLE - CONDITIONED NORMALIZING FLOWS
7.8.3.1 Importance of Domain Knowledge in Financial Time Series
7.9 CONCLUSION
8. GENERATIVE ADVERSARIAL NETWORKS
8.1 INTRODUCTION
8.2 TRAINING
8.2.1 EVALUATION
8.3 SOME THEORETICAL INSIGHT IN GANS
8.4 WHY IS GAN TRAINING HARD? IMPROVING GAN TRAINING TECHNIQUES
8.5 WASSERSTEIN GAN (WGAN)
8.5.1 GRADIENT PENALTY GAN (WGAN-GP)
8.6 EXTENDING GANS FOR TIME SERIES
9. LEVERAGING LLMS FOR SENTIMENT ANALYSIS IN TRADING
9.1 SENTIMENT ANALYSIS IN FED PRESS CONFERENCE SPEECHES USING LARGE LANGUAGE MODELS
9.2 DATA: VIDEO + MARKET PRICES
9.2.1 COLLECTING AUDIO DATA
9.3 SPEECH TO TEXT CONVERSION
9.3.1 WHISPER MODEL
9.3.1.1 Python usage
9.3.2 WHISPER ON FED SPEECH AUDIO DATA
9.3.3 AUDIO SEGMENTATION
9.4 SENTIMENT ANALYSIS
9.4.1 BERT
9.4.1.1 BERT Overview
Input/Output Representations
Input Representations
Output Representations
Pre-training objectives
9.4.1.2 Fine-Tuning BERT for Enhanced Financial Sentiment Analysis: Producing FinBERT
9.4.1.3 Using FinBERT
9.5 EXPERIMENT RESULTS
9.6 CONCLUSION
10. EFFICIENT INFERENCE
10.1 INTRODUCTION
10.2 SCALING LARGE LANGUAGE MODELS: HIGH PERFORMANCE, HIGH COMPUTATIONAL COST, AND EMERGENT ABILITIES.
10.2.1 EMERGENT ABILITIES
10.2.2 IMPACT OF MODEL SIZE
10.2.3 EFFECT OF TRAINING TIME.
10.2.4 EFFICIENT INFERENCE FOR DEEP MODELS.
10.3 MAKING FINBERT FASTER
10.3.1 KNOWLEDGE DISTILLATION
10.3.1.1 Which aspect of the teacher model to match
10.3.1.2 Response-based knowledge.
10.3.1.3 Implementation details.
10.3.2 CASE STUDY RESULTS. DISTILLING FINBERT
10.4 MODEL QUANTIZATION
10.4.1 LINEAR QUANTIZATION
10.4.1.1 Example of Linear Quantization
10.4.2 QUANTIZING AN ATTENTION LAYER IN DISTILLED FINBERT
10.4.3 EXPERIMENT RESULTS WITH LINEAR QUANTIZATION ON DISTILLEDFINBERT
10.5 CUSTOMIZING YOUR LLM: ADAPTING MODELS TO YOUR NEEDS
10.5.1 FINE-TUNING TECHNIQUES.
10.5.1.1 Traditional Fine-Tuning (FT):
10.5.1.2 Parameter-Efficient Fine-Tuning (PEFT)
BitFit
Adapters
Prompt-tuning
LoRA (Low-rank adaptation of large language models)
Efficiency of LoRA
QLoRA
10.5.1.3 Aligning Your LLM with Human Preferences
10.6 CONCLUSIONS
11. AFTERWORD
REFERENCES
APPENDICES
APPENDIX A -
A.1 RETRIEVING ADJUSTED CLOSING PRICES AND COMPUTING DAILY RETURNS
A.2 INSTALLING PYTHON
A.2.1 STEP 1: DOWNLOAD PYTHON
A.2.2 STEP 2: INSTALL PYTHON
A.2.3 STEP 3: SET UP A VIRTUAL ENVIRONMENT (OPTIONAL BUT RECOMMENDED)
A.2.4 STEP 4: INSTALL PACKAGES WITH PIP
A.2.5 STEP 5: CONSIDER AN INTEGRATED DEVELOPMENT ENVIRONMENT (IDE)
A.2.6 ADDITIONAL TIPS
A.3 PLOTTING THE RISK-FREE-RATE OVER THE YEARS
A.4 COMPUTING THE SHARPE RATIO OF SPY
A.5 MATLAB CODE FOR COMPUTING EFFICIENT FRONTIER AND FINDING THE TANGENCY PORTFOLIO
APPENDIX B -
B.1 COMPUTING NEXT-DAY’S RETURN
B.2 UPLOADING THE FAMA-FRENCH FACTORS
B.3 COMBINING FAMA-FRENCH FACTORS WITH NEXT-DAY’S RETURNS
References
Index
