ICAIF 2025 TUTORIAL

DECISION-FOCUSED LEARNING IN FINANCIAL OPTIMIZATION.

Bridging Prediction and Optimization at ICAIF 2025

Learn how Decision-Focused Learning directly aligns machine learning training objectives with the quality of downstream decisions. Through hands-on exercises with PyTorch and cvxpylayers, practice the techniques that achieve superior financial decisions.

NOV 15

Saturday, 2025

3.5 HRS

Duration

IN-PERSON

Format

Tutorial Schedule

3 Hands-on Exercises • Interactive Sessions

Session 1

Introduction & Motivation

Speaker: Yongjae Lee

Overview of Decision-Focused Learning (DFL) vs. Prediction-Focused Learning (PFL)
Motivation in financial applications: mean-variance optimization, goal-based investing
Key conceptual distinctions and intuition behind decision-aware training
Why prediction accuracy doesn't always lead to better decisions

Session 2

Background in Decision-Focused Learning

Speaker: Haeun Jeon

PFL vs DFL Pipeline: Understanding the fundamental paradigm shift from two-stage to end-to-end learning
Theoretical Considerations: Challenges of back-propagating through optimization problems, handling non-differentiable structures
[Hands-on Exercise] PyTorch Implementation:
- Building a simple DFL model from scratch
- Integrating cvxpylayers for differentiable optimization
- Observing gradient flow from decisions to predictions

Session 3

DFL in Mean-Variance Optimization

Speaker: Junhyeong Lee

Markowitz Framework Review: Mathematical formulation and parameter estimation techniques
[Hands-on Exercise] Complete DFL Pipeline Implementation:
- Neural network for return prediction
- Differentiable optimization layer integration
- End-to-end training loop implementation
- Real-time experimentation with architectures
PFL vs DFL Analysis: Live performance comparison with real market data, visualizing how DFL learns different prediction patterns

Session 4

DFL in Partial Index Tracking

Speaker: Hyunglip Bae

Partial Index Tracking: Replicating benchmark performance with limited assets
Semi-definite Relaxation: Convex relaxation method for cardinality constraints
[Hands-on Exercise] DFL in partial index tracking:
- Solving partial index tracking problem with Cvxpy
- Implementing DFL for partial index tracking problem with CvxpyLayer
- Performance comparison of PFL and DFL in partial index tracking