Foundation Models for Financial Time Series — Final Exam

This is the graded finale for Foundation Models for Financial Time Series, and it runs the full arc the course built. You started with the pretrain-and-adapt bet — Bommasani’s 2021 definition of a model trained broadly at scale and adapted to many tasks — and the two-phase amortization economics that justify it: one big fixed pretraining cost plus a cheap per-task adaptation, so a 40-fixed-plus-1-per-task curve crosses a from-scratch 10-per-task curve at roughly $5$ tasks. You learned how time-series foundation models (TimesFM, Chronos, Moirai, Lag-Llama) tokenize continuous values two ways — patching ( $512$ context $/\,32$ patch $=16$ tokens) and Chronos’s scaling-plus-quantization into a fixed vocabulary — and why you mean-scale on the context only. You walked the adaptation spectrum from zero-shot to fine-tune, met MASE ( $<1$ beats naive) and the scaling law $L(N)=L_\infty+(N_c/N)^\alpha$ whose irreducible floor $L_\infty$ flattens the curve on markets. You saw LLMs split into signal versus researcher roles and the data-snooping explosion where best-of-N Sharpe inflates with $\sqrt{2\ln N}$ . You mapped the leakage minefield that foundation models make worse — temporal contamination, benchmark contamination, point-in-time revisions — and finished with the honest evaluation: the deflated-Sharpe hurdle, purged/embargoed CV, and the end-to-end test of surviving costs, capacity, and a post-cutoff out-of-sample run. No hints are shown, each answer locks the moment you submit, and your score stays hidden until the very end.

Course Recap

Big picture

Foundation Models for Financial Time Series — the whole arc

FM for Financial Time Series
- 1 · The pretrain-and-adapt bet
  - Foundation model = broad pretraining, many-task adaptation (Bommasani 2021)
  - Amortization: big fixed cost + cheap per-task; crossover ≈ 5 tasks
  - Markets non-stationary, reflexive, low-SNR → transfer doubtful
- 2 · Time-series foundation models
  - TimesFM / Chronos / Moirai / Lag-Llama
  - Patching vs Chronos scaling+quantization; mean-scale on context only
  - Probabilistic next-patch objective; corpora mostly non-finance
- 3 · Zero/few-shot & scaling laws
  - Zero-shot = no weight updates; MASE < 1 beats naive
  - Adaptation spectrum zero → few → fine-tune; in-context learning
  - L(N)=L∞+(Nc/N)^α; noise floor L∞ flattens the curve
- 4 · LLMs as research tools & signals
  - Signal: sentiment/event extraction, decays in days
  - Researcher: hypotheses/features/code, can invent nonsense
  - Data-snooping: best-of-N Sharpe inflates with √(2 ln N)
- 5 · The leakage minefield
  - Temporal contamination: backtest only after the cutoff
  - Benchmark contamination inflates zero-shot scores
  - Point-in-time: availability time, vintage data, survivorship
- 6 · The honest evaluation
  - Deflated Sharpe + purged/embargoed CV
  - Helps: cross-asset, cold-start, data-poor regimes
  - Hype: liquid arbitraged low-SNR return forecasting

Six lessons, one thesis: borrowed scale is powerful, but markets punish naive transfer.

One run, one shot

This is a graded, irreversible exam. There are 25 questions, shown one at a time. The instant you submit a question it locks — there is no Back button, no retry, and no Restart. A wrong answer simply fails that question and the exam moves on; you cannot revisit it. Your running score is hidden until the final screen. The pass mark is 70%. Some questions accept more than one correct option — read every option before you commit, because once you submit you own the answer.

Question 1 of 25

The 2021 Stanford report by Bommasani and colleagues introduced the term 'foundation model.' Which definition matches the one used in that work?

A model pretrained on broad data at scale that can then be adapted to a wide range of downstream tasks.A model trained end-to-end on a single narrow task with no reuse across problems.A statistical model whose foundations are derived analytically rather than learned from data.Any neural network with more than one billion parameters, regardless of how it is trained.

Select an answer to continue.

Where this leaves you on the quant ladder

Pass or fail, you now hold the rare judgment that this whole field is short on: knowing when borrowed scale is a genuine head start and when it is a treadmill. You can read a foundation model’s amortization math, see through patching and quantization to the tokenizer underneath, demand context-only scaling, and tell a real zero-shot forecast from a memory test. You can wield an LLM as both signal and researcher while deflating its best-of-N Sharpe, and you can walk the leakage minefield — temporal, benchmark, and point-in-time — without stepping on a mine. Most of all, you can run the honest end-to-end test: costs, capacity, deflation, and a strictly post-cutoff out-of-sample run. That is exactly the discipline that lets you deploy foundation models where they truly help — cross-asset, cold-start, data-poor regimes — and walk away from the hype everywhere else.