[KongchangAI]
ModelToto

Toto 2.0

A time-series foundation model introduced by Datadog, aimed at scenarios such as observability metrics, supporting zero-shot and multivariate forecasting.

Timeline (last 90 days)

Nov 10

In zero-shot next-day direction prediction on 24 large U.S. stocks and ETFs, the direction accuracy of TimesFM 3.0, Chronos-2 and Toto 2.0 ranged from 49.9% to 51.1%, comparable to random guessing.

Unverified50%
Nov 10

On the same stock dataset, a naive strategy that always predicts an increase achieved next-day direction accuracy of 52.2%, higher than the zero-shot performance of the three time-series foundation models.

Unverified50%
Nov 10

When simulated random walk data constructed to contain no exploitable predictive signal was fed into the three time-series foundation models, their direction accuracy was roughly the same as on real stock price data.

Unverified50%
Nov 10

In a simulated trading test (initial capital of $10,000, with a 0.05% trading cost per switch), the long-short switching strategy based on time-series model signals ended with final assets between $10,893 and $12,181, while a buy-and-hold strategy over the same period ended with $15,036.

Unverified50%
Nov 10

TimesFM 3.0 is developed by Google, Chronos-2 by Amazon, and Toto 2.0 by Datadog.

Unverified50%
Nov 10

On 5-minute candlestick data, a simple one-line mean-reversion heuristic strategy outperformed the zero-shot direction predictions of the three time-series foundation models.

Unverified50%

All Facts (6)

Source Articles