Time-series foundation models for finance

Does giving a forecasting model more series make it more accurate?

A study of Amazon's Chronos-2, a pretrained time-series transformer, on Magnificent-7 equities and U.S. Treasury rates. We compare univariate forecasts against multivariate ones — and let you run the real forecasts yourself.

Explore forecasts →See the results Read the paper ↗

By Sanjiv Das, Tarang Goyal & Mohini Yadav

−60%

Rates forecast error

MV vs UV MAPE (4.97% vs 12.33%)

−16%

Equity forecast error

MV vs UV MAPE (7.06% vs 8.44%)

Series, jointly modeled

7 equities + 10 Treasury maturities

Explore

Test it yourself

Pick a panel, a series, and a horizon, then scrub through a decade of monthly forecasts. Watch the multivariate and univariate errors diverge over time, and see the win-rate across the whole panel — all computed live from real Chronos-2 outputs.

Loading real Chronos-2 forecasts…

Coming soonBring your own series

Paste any time series and get a live Chronos-2 forecast on demand. This runs the model at request time on a GPU endpoint — being wired up next.

Method

Univariate vs multivariate, one foundation model

Univariate

Forecast each series from its own history alone. One input series in, one forecast out — the classic baseline.

Multivariate

Feed all related series together (e.g. the whole yield curve) and let Chronos-2 attend across them. Same model, richer context.

The setup

• Input window n = 252 trading days (rolling)
• Horizons m = 21 and 63 days (1 & 3 months)
• Panels: Mag-7 equities, 10 Treasury maturities, and both combined
• Metrics: MAPE (primary) and RMSE on realized values
• Zero-shot: Chronos-2 is not fine-tuned on these series

Research questions

1.Do multivariate (MV) inputs beat univariate (UV) ones when a foundation model forecasts both?
2.Is the MV advantage larger for stocks or for interest rates?
3.Does forecasting stocks and rates jointly — a small “world” model — help or hurt?
4.Are the gains an artifact of pre-training leakage, or real?

Findings

Multivariate inputs consistently beat univariate ones

Across the full rolling evaluation, feeding Chronos-2 related series together lowers error for every rate and every stock — with the largest gains in interest rates. These are the paper's aggregate numbers.

Rates

−60% MAPE

Multivariate4.97%

Univariate12.33%

Mean MAPE across all windows, horizons, and 10 maturities, 2000–2025.

Stocks

−16% MAPE

Multivariate7.06%

Univariate8.44%

Mean MAPE across all windows, horizons, and 7 equities, 2000–2025.

Treasury rates — MAPE by maturity

3-Month

23.55%

31.86%

6-Month

8.33%

17.70%

1-Year

5.72%

13.26%

2-Year

3.34%

12.70%

3-Year

2.43%

11.81%

5-Year

1.71%

9.72%

7-Year

1.34%

8.27%

10-Year

1.14%

7.03%

20-Year

0.99%

5.72%

30-Year

1.13%

5.20%

Equities — MAPE by ticker

Apple

5.99%

7.20%

Amazon

5.67%

7.07%

Alphabet

4.78%

6.35%

Microsoft

3.68%

4.66%

Netflix

10.40%

11.45%

NVIDIA

9.06%

11.07%

Tesla

10.98%

12.42%

Multivariate Univariate(lower is better)

But a bigger “world” model doesn't help

Forecasting stocks and rates jointly — a combined 17-series panel — slightly degrades accuracy versus modeling each asset class on its own. Cross-domain context adds noise, not signal. MAPE, individual panel vs combined:

3-Month23.04% → 23.92%

2-Year3.66% → 3.92%

10-Year1.28% → 1.26%

NVDA7.43% → 7.61%

MSFT3.05% → 3.18%

TSLA10.90% → 11.14%

Robustness

Is this just training leakage?

A fair worry: maybe Chronos-2 already saw these series in pre-training. Two checks argue against it. If the gains were leakage, (1) MV and UV would look alike — they don't, and (2) forecasts before the ~2023 training cutoff would beat those after. Instead, post-2023 forecasts are more accurate, on data the model could not have seen. Suggestive, not conclusive — but the multivariate advantage is not an artifact.