Foundation Models for Financial Time Series
The rest of AI stopped training one model per task and started pretraining a giant model, then adapting it. This course drags that playbook into markets — time-series foundation models, zero-shot forecasting, LLM-driven research — and refuses to let it off easy: foundation models are trained on near-stationary worlds, and markets are adversarial and non-stationary, so a leaderboard win is almost never a tradeable edge.
The pretrain-and-adapt playbook walks into markets — large time-series foundation models (the TimesFM / Chronos / Moirai / Lag-Llama lineage), how they tokenize continuous returns, zero-shot and few-shot forecasting and whether scaling laws even hold on financial data, LLMs as research tools and as multimodal price-plus-text signals, the leakage minefield of pretraining on the future, and an honest, deflated-Sharpe audit of whether any of it survives an efficient, non-stationary market.
The rest of artificial intelligence went through a quiet revolution while quants were busy tuning one gradient-boosted tree at a time. The old recipe — collect a labeled dataset for your task, train a model from scratch, repeat for the next task — got replaced by a two-step playbook so effective it now defines the field: pretrain one enormous model on a mountain of broad data, then adapt it cheaply to hundreds of downstream tasks. A single pretrained network underwrites translation, summarization, code, and chat; a single pretrained vision model underwrites detection, segmentation, and captioning. The model that does this is a foundation model, and the bet behind it is that a representation learned once, on enough data, transfers.
This course drags that playbook into financial time series — and then refuses to let it off easy. We assume you finished Deep Learning for Market Data (you know why deep nets struggle on low-signal returns, and you can defend a result with a deflated Sharpe and a purged, embargoed split) and Generative Models for Synthetic Market Data (you know the stylized facts a market series must reproduce and how to evaluate a generator without fooling yourself). Here we add the frontier layer: stop training a model per task, pretrain a foundation model, and adapt it. The whole arc is a single sustained argument — here is the machinery, here is exactly where it could help, and here is the brutal reason it usually doesn’t.
- The pretrain-and-adapt bet — what a foundation model actually is, why the two-phase economics (one big fixed pretraining cost, many cheap adaptations) conquered language, vision and weather, and the central tension the moment you point it at markets: amortization only pays off if the pretrained representation transfers, and markets are adversarial, non-stationary and brutally low-signal — transfer is exactly what’s in doubt.
- Time-series foundation models — the TimesFM / Chronos / Moirai / Lag-Llama lineage, the decoder-only transformer over patches, how you turn a continuous stream of returns into tokens a transformer can read (patching, mean-scaling, Chronos-style quantization into a vocabulary), and what oceans of data these models are pretrained on.
- Zero-shot & few-shot forecasting, and scaling laws — forecasting a series the model has never seen with no fitting, the spectrum from zero-shot to few-shot to full fine-tuning, in-context learning, and the question that decides everything: do the neural scaling laws that hold on text and images hold on financial data, or does the noise floor cap you long before more data or parameters help?
- LLMs as research tools and as signals — large language models pointed at finance two ways: as a signal (sentiment and event extraction from filings and news, multimodal price-plus-text models) and as a researcher (LLM agents that generate hypotheses and write strategy code), plus the data-snooping trap that an automated idea generator makes terrifyingly easy.
- The leakage minefield — the failure mode foundation models make worse than anything before them: lookahead in the pretraining corpus. A model pretrained on text and prices through last year has, in a real sense, read the future relative to your backtest — pretraining-data contamination, point-in-time discipline, and why the gap between a published benchmark win and a tradeable strategy is mostly leakage.
- The honest evaluation — carrying the deflated-Sharpe and purged-CV creed all the way: foundation models are trained on near-stationary domains (language, weather, demand) and markets are adversarial and non-stationary, so benchmark wins rarely translate to PnL. The map of where pretraining genuinely helps — cross-asset transfer, cold-start, data-poor regimes — versus where it is pure hype.
By the end you’ll be able to explain how a time-series foundation model is built and adapted, wire an LLM in as a signal or a research assistant without leaking the future into your backtest, and — most valuably — judge, with a deflated out-of-sample number rather than a glittering leaderboard, whether the pretrain-and-adapt revolution earns its keep on your market or just borrows credibility from someone else’s. A graded final exam runs the whole discipline back at you at the end, one locked question at a time.
In this topic
- 1 The Pretrain-and-Adapt Bet What a foundation model actually is, the amortization economics that make 'pretrain once, adapt cheaply' worth it, why the recipe conquered language, vision and weather — and the central, unresolved tension when you point it at non-stationary, adversarial, low-signal financial markets. 17 min
- 2 Time-Series Foundation Models: TimesFM, Chronos, Moirai & Lag-Llama How a forecasting model pretrained once on millions of series replaces the one-model-per-series ritual — tokenizing a continuous series by patching or by scaling-and-quantizing, the autoregressive probabilistic pretraining objective, and the TimesFM / Chronos / Moirai / Lag-Llama lineage, with a frank look at why their corpora barely resemble markets. 19 min
- 3 Zero-Shot, Few-Shot & Scaling Laws Zero-shot forecasting with no weight updates, the MASE benchmark that decides whether you beat the naive baseline, the zero-shot → few-shot → full fine-tune adaptation spectrum, in-context learning as a tiny inference-time training set, and the neural scaling laws — plus the irreducible-noise-floor argument for why they flatten at a useless level on near-efficient market returns. 18 min
- 4 LLMs as Research Tools & as Signals The two jobs an LLM can hold in a quant stack — as a signal that turns text into a return-predicting number, and as a researcher that proposes hypotheses, features, and code — with sentiment-and-event extraction, multimodal price-plus-text fusion, the AI-quant idea loop, and the data-snooping explosion that an automated idea firehose detonates. 18 min
- 5 The Leakage Minefield Why foundation models make data leakage worse than any model before — pretraining contamination, the knowledge-cutoff trap, benchmark overlap, point-in-time and survivorship leaks — plus a concrete checklist for when a zero-shot foundation-model backtest can actually be trusted. 18 min
- 6 The Honest Evaluation: Does Any of It Survive? The finale: re-arm the deflated Sharpe and purged-embargoed cross-validation, see why near-stationary pretraining benchmarks don't transfer to adversarial markets, map exactly where pretraining genuinely helps (cross-asset transfer, cold-start, data-poor regimes) versus where it's pure hype, and run the end-to-end test that turns a benchmark win into actual money. 18 min
- 7 Foundation Models for Financial Time Series — Final Exam The graded final exam for Foundation Models for Financial Time Series: the pretrain-and-adapt bet and its amortization economics, time-series tokenization via patching and Chronos scaling-quantization, zero/few-shot transfer and scaling laws with a noise floor, LLMs as signals versus researchers and the data-snooping explosion, the foundation-model leakage minefield, and the honest evaluation that survives deflation, purged CV, and post-cutoff out-of-sample tests. 20 min
Mark course as finished
Done with every lesson? Lock it in — your progress is saved on this device.