Model Watch
Short verdicts on time-series and financial ML models and papers.
A summary of a published model or paper. Not a recommendation to use it for trading, and not financial, investment, legal or tax advice. Summaries can be inaccurate or out of date and are provided "as is". Check the model's license before commercial use. Terms & Disclaimer
View Archive →Tail-GAN: Learning to Simulate Tail Risk Scenarios
GAN-based approach for generating high-dimensional multi-asset price scenarios that preserve the tail risk (Value-at-Risk and Expected Shortfall) of a set of trading strategies, for estimating portfolio loss distributions.
Novel data-driven scenario generator built specifically for tail risk. In the authors' experiments it captures VaR and Expected Shortfall better than earlier data-driven generators, including a Wasserstein GAN baseline, by exploiting the joint elicitability of VaR and ES. Training on PCA eigenportfolios lets it scale to more assets (shown on 20 simulated assets). Strong for tail-risk scenario generation and risk-management research.
Chronos / Chronos-2 (Amazon Science)
Pretrained time-series forecasting models. Chronos-2 (120M params) supports zero-shot univariate, multivariate, and covariate forecasting. Family spans 8M to 710M parameters. (Paper with Apache-2.0 code)
The README reports state-of-the-art zero-shot results for Chronos-2 across multiple benchmarks, including fev-bench and GIFT-Eval. Apache-2.0 licensed. Install with pip install chronos-forecasting. The README claims Chronos-Bolt is up to 250x faster than original Chronos models. Strong option for practical forecasting tasks with flexible model sizes.
TimesFM: A Decoder-Only Foundation Model for Time-Series Forecasting
Google Research pretrained decoder-only time-series foundation model (TimesFM) for zero-shot forecasting; ICML 2024 paper. Version 3.0 adds multivariate and covariate support. (Paper with Apache-2.0 code; v1.0–2.5 weights are Apache-2.0, v3.0 weights are non-commercial, non-production)
Useful zero-shot baseline for multi-domain series before training a bespoke forecaster. TimesFM 3.0 claims rank #1 on fev-bench, TIME Benchmark, and GIFT-Eval. Available in BigQuery ML and Vertex AI Model Garden for commercial use. Open version not officially supported by Google; v3.0 weights require non-commercial license (production use needs Google Cloud).
FinGPT: Open-Source Financial LLMs
Open-source large language model framework for financial NLP using data-centric approach with automatic data curation and low-rank adaptation (LoRA) for efficient fine-tuning.
Framework code released under the MIT license for fine-tuning LLMs on financial text with LoRA. Released adapters cover sentiment, multi-task and forecasting and use base models including Llama 2 and ChatGLM2; the licenses for those base models also apply. In the README's weighted-F1 table for FPB, FiQA-SA and TFNS, the authors report higher scores for their Llama2-13B sentiment adapter (FinGPT v3.3) than for GPT-4 on all three.
Temporal Fusion Transformer for Interpretable Multi-horizon Forecasting
Attention-based architecture for multi-horizon time series forecasting with interpretable outputs.
Strong performance on financial time series. Attention mechanism helps identify important features. Computationally expensive but worth it for complex forecasting tasks.
Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting
Efficient transformer architecture using ProbSparse self-attention for long-sequence time series forecasting with O(L log L) complexity.
Handles long sequences better than vanilla transformers. Useful for macro forecasting and long-term trend prediction. Implementation complexity is moderate.
N-BEATS: Neural Basis Expansion Analysis for Time Series
Deep neural architecture specifically designed for univariate time series forecasting with backward and forward residual links.
Pure deep learning approach with no feature engineering. Impressive results on M4 competition. Good baseline for financial forecasting.
Empirical Asset Pricing via Machine Learning
Comparative analysis of machine learning methods for measuring equity risk premiums (expected stock returns), using 94 stock characteristics, their interactions with eight aggregate time-series variables, and 74 industry dummies.
Comprehensive empirical study of ML in asset pricing. Trees and neural networks perform best by capturing nonlinear predictor interactions. Highlights the value and limitations of ML for measuring risk premia.
Deep Hedging: Hedging Derivatives with Reinforcement Learning
Deep reinforcement learning framework for hedging derivatives under realistic market frictions including transaction costs, market impact, and risk limits.
Framework from the paper (Quantitative Finance, 2019; arXiv 2018) for hedging a portfolio of derivatives under market frictions such as transaction costs, market impact, liquidity constraints or risk limits, using deep reinforcement learning. In the abstract, the authors report that, in a synthetic market driven by the Heston model with transaction costs, they outperform the standard "complete market" solution.
Advances in Financial Machine Learning (Book)
Marcos López de Prado's foundational book on financial ML techniques. Published by Wiley. Paid resource.
Essential reading. Covers cross-validation, feature importance, and meta-labeling for financial data. Addresses practical pitfalls that academics often miss. This is a copyrighted book, not open-source software.