Forecasting with Artificial Intelligence: Theory and Applications

Foreword
Preface
Acknowledgments
Contents
Editors and Contributors
About the Editors
Contributors
Abbreviations
List of Figures
List of Tables
Part I Artificial Intelligence: Present and Future
1 Human Intelligence (HI) Versus Artificial Intelligence (AI) and Intelligence Augmentation (IA)
Introduction
Defining Intelligence
The Distinctive Nature of AI and HI
The Evolution of HI and Its Complementarity with AI
AI: Current Capabilities and Future Challenges
The Future of AI: Uncertainties and Challenges
AVs: Achievements, Limitations, and Future Prospects
The Current Situation
The Four AI/HI Scenarios
Smarter, Faster Hedgehogs
An Army of Specialized Hedgehogs
Fully Integrated HI and AI
More Efficient Ways of Communicating with Computers
Direct Brain to Computer Interface (BCI)
Non-Invasive Forms of Interfaces
Invasive Forms of Interfaces
Brain-To-Brain Interfaces (BBI)
The Future World of Transhumanism
Runaway AGI
Conclusions
References
2 Expecting the Future: How AI’s Potential Performance Will Shape Current Behavior
Introduction
The Black Box of Machine Learning
Decreased Efforts
Educational Choices
Political Job Protection
Greater Desire of Governments to Borrow
Economies of Scale
Trade Blocks
What Happens If People Expect Money to Become Worthless
Animal Welfare
Conclusion
References
Part II The Status of Machine Learning Methods for Time Series and New Product Forecasting
3 Time Series Forecasting with Statistical, Machine Learning, and Deep Learning Methods: Past, Present, and Future
Introduction
Statistical Methods
Machine Learning Methods
Deep Learning Methods
Discussion and Future Steps
References
4 Machine Learning for New Product Forecasting
Introduction
Categories of New Products and Data Availability
Data Availability
There is Data on Similar Products that Have Been Adopted in the Past
The Product is Truly New
Highlights of the Main Forecasting Techniques Used for New Products
“Traditional” Methods
Expert Opinion
The Delphi Method
Consensus of Sales Team
Customers Surveys
Drivers and Constraints
Diffusion Models (S-Shaped Curves)
Machine Learning Methods for Predicting New Product Demand
Gradient Boosted Trees (GBT)
Artificial Neural Network (ANN)
Structuring the New Product Forecasting Problem
Potential Machine Learning Approaches for Truly New-to-the-World Products
A Review of Four Case Studies of Machine Learning for New Product Forecasting
Forecasting Demand Profiles of New Products
Description
Test Setup
Results
Feature Importance and Comparable Products
Summary and Conclusion
Fashion Retail: Forecasting Demand for New Items
Methodology
Experiment and Results
Conclusion and Future Directions
New Product Forecasting Using Deep Learning - A Unique Way
Methodology and Results
Summary and Conclusion
Forecasting New Product Demand Using Machine Learning
Methods
Results
Summary and Conclusion
Summary and Lessons from the Four Case Studies
Summary and Conclusions
References
Part III Global Forecasting Models
5 Forecasting with Big Data Using Global Forecasting Models
Big Data in Time Series Forecasting
Global Forecasting Models
A Brief History of Global Models
Model Complexity in Global Models
Global Forecasting Model Frameworks
Recent Advances in Global Models
Conclusions and Future Work
References
6 How to Leverage Data for Time Series Forecasting with Artificial Intelligence Models: Illustrations and Guidelines for Cross-Learning
Introduction
Background and Definitions
Common Terminology and Building Blocks
Locality and Globality
Approximation Capacity of Cross-Learned Models
Statistical Trade-Offs
Illustration of Cross-Learning Approximation of Specific Time Series Processes
Well-Known Families of Time Series Processes Have an Equivalent Parameterization via Local Autoregression
Two Different Linear Autoregressive Predictive Functions Can Be Represented by a Single Linear Autoregressive with More Lags
A Single Nonlinear Autoregression Can Represent an Arbitrary Number of Different Linear Autoregressions
Cross-Learning for Processes with Noise or Perturbations
Explicit Example of Transfer Learning
Conclusions
References
7 Handling Concept Drift in Global Time Series Forecasting
Introduction
Problem Statement: Concept Drift Types
Related Work
Global Time Series Forecasting
Transfer Learning
Concept Drift Handling
Methodology
Series Weighted Methods
Our Proposed Concept Drift Handling Methods
Error Contribution Weighting
Gradient Descent Weighting
Experimental Framework
Datasets
Error Metrics
Evaluation
Benchmarks
Statistical Testing of the Results
Results and Discussion
Main Accuracy Results
Statistical Testing Results
Further Insights
Conclusions and Future Research
References
8 Neural Network Ensembles for Univariate Time Series Forecasting
Introduction
Experimental Design
Initialization Seed
Loss Function
Input Layer Size
Empirical Evaluation
Time Series Data
Forecasting Performance Measures
Base Forecasting Neural Network
Results and Discussion
Initialization Seed
Loss Function
Input Layer Size
Final Ensemble
Conclusions
References
Part IV Meta-Learning and Feature-Based Forecasting
9 Large-Scale Time Series Forecasting with Meta-Learning
Introduction
A Brief Review of the Meta-Learning in Time Series Forecasting
Key Choices in Designing an Effective Meta-Learning Framework
Selection of Base Forecasters
Feature Extracting
Meta-Learners
A Python Library for Time Series Forecasting with Meta-Learning
Raw Data Class
Metadata Class
Meta-Learner
The Meta-Combination Learner with Encapsulated Features Extractor (‘MetaComb’)
The Meta-Combination Learner with Unsupervised Hand-Selected Features (‘MetaCombFeat’)
The Model Selection Learner Targets Identifying the Best Base Forecaster (‘MetaSelection’)
The Metaloss Prediction Learner (‘MetaLoss’)
Experimental Evaluation
Datasets
Predictive Performance
Conclusions and Future Directions
References
10 Forecasting Large Collections of Time Series: Feature-Based Methods
Introduction
Data Generation
Diverse Time Series Generation
Time Series Generation with Target Features
Feature Extraction
Time Series Features
Automation of Feature Extraction
Time Series Imaging
Forecast Diversity
Automatic Feature Selection
Forecast Trimming for Combination
Accuracy, Robustness, and Diversity
Forecast Trimming
Some Practical Forecasting Issues
Intermittent Demand
Uncertainty Estimation
Conclusion
References
Part V Special Applications
11 Deep Learning Based Forecasting: A Case Study from the Online Fashion Industry
Introduction
Data for Forecasting at Zalando: An Overview
Sales to Demand Translation
Description of Covariates
Demand Forecast Model
Problem Formalization
Model Architecture
Input Preparation
Encoder
Positional Encoding
Padding and Masking
Decoder
Monotonic Demand Layer
Near and Far Future Forecasts
Training
Prediction
Empirical Results
Accuracy Metrics
Model Benchmarking
On the Benefits of Transformer-based Forecasting: First Results
Related Work
Practical Considerations
Conclusion
References
12 The Intersection of Machine Learning with Forecasting and Optimisation: Theory and Applications
Introduction
Predict and Optimise
Problem Setting
Methodologies and Applications
Discussion and Future Works
Predicting the Solutions of Optimisation Problems
Problem Setting
Methodologies and Applications
Discussion and Future Works
Conclusion
References
13 Enhanced Forecasting with LSTVAR-ANN Hybrid Model: Application in Monetary Policy and Inflation Forecasting
Introduction
LSTVAR-ANN Hybrid Model
LSTVAR Model
LSTAR-ANN Hybrid Model
Research Architecture Specification
3D Impulse Response Function
Monetary Policy Model with Regimes
Data
Results
Unobserved Factors Extraction
Importance on CPI Prediction
Transition Function Output
Response of CPI on Interest Rate Shock
Forecasting
Pre-COVID-19 Period
COVID-19 and Post-COVID-19 Period
Conclusion
Appendix
References
14 The FVA Framework for Evaluating Forecasting Performance
What is FVA?
Why Use FVA?
Applying the Scientific Method to Forecasting
FVA Analysis: Step-by-Step
Mapping the Process
Collecting the Data
Analyzing the Process
Reporting the Results
Interpreting the Results
FVA Challenges and Extensions
Compete or Combine?
Stochastic Value Added (SVA): FVA for Probabilistic Forecasts
Summary
References
Author Index
Subject Index