Do LLMs really improve time series forecasting?

View profile for Alexander Denev

Turnleaf Analytics (Forecasting Macro and Inflation with Machine Learning and Alternative Data)

🔮 Are LLMs really the future of time series forecasting? There’s growing hype around using large language models (LLMs) for forecasting—but does the evidence support it? 📉 For time series (especially in trading), models must respect causality—only using information available up to each point in time. Without this, you risk forward bias and information leakage. But LLMs, by design, don't embed a true notion of temporal structure. Recent studies like “Are Language Models Actually Useful for Time Series Forecasting?” (NeurIPS 2024) support this point. In other words: if you can identify the right features at the right time, even a linear model can outperform a black-box LLM trained on irrelevant inputs. This is especially true in markets, where the signal is sparse and noise is high. Overfitting to irrelevant patterns is a real risk. Are you using LLMs for time series? What’s working—and what’s not? #timeseries #forecasting #llms #machinelearning #financemodeling #causality #datascience #quantitativefinance

Borislav Vladimirov

Brevan Howard Asset Management

2mo

Language, images etc are high complexity martingales, (fixed variance process), where compute power and large training sets deliver reasonably good approximations. Markets, traffic, consumption and political preferences are causal random walks. LLMs are not suitable.

Saeed Amen

Co-founder at Turnleaf Analytics / Macro forecasting with ML

2mo

These are excellent points! I think part of the reasons LLMs have been explored for time series is because of their success in "forecasting" language/the next token in a sentence. The difficulty is (as you point out) markets are somewhat different, where the signal is sparse with lots of noise and ultimately markets are less stationary than language.

Marcos Lopez de Prado

Global Head - Quantitative R&D at ABU DHABI INVESTMENT AUTHORITY (ADIA), Professor of Practice at CORNELL UNIVERSITY

1mo

Excellent observations, Alexander Denev . Overfitting remains the main issue.

Marco Jean Aboav, PhD

Building neuro-symbolic financial super intelligence @Etna Research

2mo

LLMs and time series is not a match made in heaven. I am surprised that so many people waste time on this topic.

Rachid Lassoued

Global Head Financial Engineering (BVAL Derivatives & MARS Valuation) & MARS Risk (Market Risk, Climate Risk, Credit Risk, XVA, Margin Analytics & Risk APIs). Opinions are my own.

2mo

Neural nets (underpinning the LMMs for text and Diffusion Models for images) work pretty well in situations when the underlying data is high signal to noise ratio like for pixels in the case of images. Financial time series datasets are notoriously low signal to noise ratio. That’s all anyone ever needs to know on that topic 😉.

Aleksander Molak

Causal Modeling: Training for Start-up & Corporate Teams || Author of "Causal Inference & Discovery in Python" || Host at CausalBanditsPodcast.com || Control For Your Confounders Before They Control You

2mo

Great points, Alexander Denev No, I don't use LLMs for time series tasks. All high quality benchmarks I know show that they are computationally expensive, yet practically not very useful models for most of the tasks in this area I've been interested in.

Kriti Doneria

7.5 years of Applied AI across CPG,Supply chain,BFSI and retail. Views are personal.

2mo

Using LLM for everything needs to stop. They are NOT built for prediction use cases

Darko Medin

AI Expert, Machine Reasoning pioneer. Biostatistician. Data Science Expert. Advisor to Biotech companies and institutions around the World on Artificial intelligence. Author and Educator. Maker of BioAIworks.

2mo

Simple answer, no LLMs have no chance of becoming the future of Time Series Forecasting...

Chris Tinker

Founding partner - Libra Investment Services

2mo

It is more than just a time series issue. Financial Markets are open, complex adaptive Systems with both Chaotic and fractal stochastic components. Where AI can provide insights is in their ability to work in higher dimensional space with DDIM type models and de-noised data. As ever, it requires a deep domain knowledge to properly execute and avoid catastrophic hallucinations 

See more comments

To view or add a comment, sign in

Explore content categories