2020 → 2025: Why I’m rebuilding Machine Learning for Trading (3rd edition) The 2nd edition shipped in 2020. Since then, the practical toolkit has changed—and so has the bar for what “useful” means in real trading research. 🧵👇

Just released my next-generation agent session viewer with analytics dashboard (Go + Svelte). A much evolved version of wesm/agent-session-viewer which is now deprecated: https://t.co/D4BDX5cfX7

If you’re using GenAI in finance, what’s the most challenging part to get right in a way you actually trust? Evaluation? Data quality? Hallucinations? Workflow integration? #MachineLearning #GenerativeAI #AlgorithmicTrading #QuantitativeFinance #Polars

There’s also a new web-based primer covering foundational ML/AI and finance, so the book can stay focused on practitioner workflows and real decision-making. Preview site drops in the next few days. Full site + 5 OSS libraries in early January.

GenAI for finance goes beyond embeddings: Transformers, LLM workflows, RAG, and knowledge-graph-backed systems for filings, research, and market narratives—designed to be testable and auditable.

The 3rd edition is a rebuild of the workflow for end-2025. Major additions include: • Generative AI & agentic AI for finance • Causal discovery & uncertainty quantification • Live trading (beyond the notebook) • A faster, Polars-first research stack

I’m deeply grateful to the many readers who’ve reached out over the years with feedback, bug reports, and ideas. That community is why the project grew to 16K+ GitHub stars and ~400 Amazon reviews (4.5/5). It’s also why this isn’t a superficial update.

@alejandroll10 @furongh @TalKachman That's the full workshop. Strong year for: - Diffusion models in portfolio optimization - Domain-specific foundation models - Synthetic data generation - Agent architecttic thinking Full thread (part 1): https://t.co/nI6EmVCpi5 Workshop page:

@alejandroll10 @furongh @TalKachman Papers: TS-Agent (agentic workflows): https://t.co/Xp0Asoax05 Financial TSFMs study:

@alejandroll10 @furongh @TalKachman Hao Ni from UCL presented two complementary approaches: 1. TS-Agent: AutoML + agentic reasoning for iterative model design. Transparent, auditable workflows that beat pure AutoML baselines. 2. Financial TSFMs: Foundation models trained on financial + synthetic data show real

@alejandroll10 @furongh @TalKachman Paper:

@alejandroll10 @furongh @TalKachman on Game of Thought—iterative LLM reasoning in game-theoretic settings. Can LLMs reason strategically against other agents? Results show iterative reasoning improves performance and reduces exploitability across multiple games. Relevant if you're building multi-agent

@alejandroll10 Prof. Furong Huang @furongh on synthetic financial data (work from Capital One). The challenge: financial AI can't access real transaction data, and free-form LLMs hallucinate when modeling money. PersonaLedger: rule-regulated world model where LLMs generate persona-driven

@alejandroll10 Paper: https://t.co/pGZJe1Na71 With minimal fine-tuning, matches or beats general foundation models on returns forecasting, volatility prediction, and consumer nowcasting. 👇

@alejandroll10 Xueying Ding (Bloomberg/CMU) on why general time-series foundation models fail on financial data. Problem: limited financial data in pre-training + negative transfer from other domains. Solution: Delphyne—a TSFM pre-trained specifically for financial time series. Handles noise,

@alejandroll10 Paper: https://t.co/WvmwOu4lCG The key insight: capture cross-asset dependencies while conditioning on asset-specific factors. More flexible than shrinkage estimators, less prone to overfitting than pure ML. 👇

@alejandroll10 Best Paper Award: Xuefeng Gao's conditional diffusion model for portfolio optimization. Instead of estimating returns directly, it learns the cross-sectional distribution of next-day returns conditioned on asset factors. Generated samples feed daily mean-variance optimization.

Alejandro Lopez-Lira (@alejandroll10) on a problem that doesn't get enough attention: LLM memorization in financial forecasting. LLMs recall historical economic values with high accuracy—even when told not to. After the knowledge cutoff, this "skill" vanishes. Implication: any

Part 2 of NeurIPS Gen AI in Finance notes. This batch included the Best Paper winner, along with practical insights on synthetic data, time-series foundation models, and why LLMs can't be trusted for economic forecasting. Workshop: https://t.co/O9iWM8lCE7 #NeurIPS2025

@jacob_choi_ @LinqAlpha @alejandroll10 That's the first half. More talks coming in part 2. Workshop page if you want to dig in yourself: