Note · October 2026

Timestral with a ruler and a pencil

Mistral AI's new forecasting models, all four sizes, on series a person could forecast by hand, scored against the best forecast possible.

Dorian Joubaud

Six things we saw, each on one of the series we drew.

The best possible forecast

Each task draws hundreds of series from a known process, so the best forecast anyone could make is known too. Its 80 % band is the dashed pair on every plot. A model's score is its quantile loss divided by that of the best possible forecast: 1 is perfect, 2 is twice the unavoidable error.

Closest to the best possible

Timestral 209M beats Chronos-2 on 53 of the 76 tasks and TimesFM 2.5 on 58. Timestral 28M, a quarter of Chronos-2's size, is on par with it.

Lines: Timestral is sure of them

Timestral standardises each patch with the mean and spread of everything before it. In that frame a line looks flat, whatever its slope, which may be why lines come easy to it. We did not test this. With 128 steps of history, Chronos-2's median keeps the line too, but its band is twelve times wider than Timestral's and sags toward the past.

Crossing zero: a default clip

By default, Timestral clamps at zero the forecast of any series whose past is non-negative. That suits demand or energy, not a balance or a net flow. TimesFM 2.5 does the same with its recommended setting. With clip_nonnegative=False, Timestral follows the line.

Exponentials: forecast the logarithm

Every model's error grows with the growth rate. At e³ over the context, the median of Chronos-2 and TimesFM 2.5 ends below the truth on every series, Timestral's on 9 %. Forecasting the logarithm, then exponentiating the quantiles, turns the curve into a line: Timestral's error then stays between 1.1 and 2.1 times the best possible.

After 1 024 steps, the band collapses

Timestral forecasts 1 024 steps in one pass. Beyond, it appends that block's median to the context and goes on. A median is smooth, so the model then sees a calm recent past, and its band narrows. That is our reading, not a test.

TimesFM 2.5 shows the same thing, sooner. Its recommended quantile head stops at 1 024 steps, and beyond, its base head feeds the median back every 128 steps. Its band then collapses at step 129: it covers 79 % of a random walk over the first 128 steps, 3 % over the rest of the first block, and almost none after. Chronos-2, which rolls out from quantile paths, keeps its band. Both models that feed back a median lose their band at their first fed-back step, which points to the decoding rather than the network, though we tried no other decoder.

On a line, Timestral curves up after the first block, and Chronos-2 bends down within it: their medians end above and below the truth on every series.

Breaks, gaps and slow cycles

Timestral follows a recent level shift and crosses a gap at little cost. With less than two periods of a slow cycle in its context, TimesFM 2.5 does better. A chirp, whose period keeps shrinking, defeats every model.

A leading series: ignored, unless shifted

Here the covariate is the target itself 96 steps ahead, given over the past: its last 96 values are the whole future. Neither model uses it, the blind spot Layer 6 found in Chronos-2, Moirai and Toto (Ghoroghchian et al., 2026). Shifted back by its lead, it becomes a covariate known in advance, which both models use.

A covariate unrelated to the target changes Timestral's error by 1 to 7 % and raises Chronos-2's by 12 to 29 %.

Every task

Each card shows the first series of a task, with the errors of Timestral 209M and Chronos-2 relative to the best possible over all its series. Pick one to see it here with the other models.

Findings

• Closest to the best possible. Timestral 209M's error is 1.54 times the unavoidable one on average, against 1.86 for Chronos-2 and 2.02 for TimesFM 2.5. The 28M already matches Chronos-2.
• Lines are easy. 1.2 to 1.5 times the best possible at every size, with a far tighter band than Chronos-2 on a short history.
• Crossing zero needs clip_nonnegative=False. The default clip holds a falling line at zero, 14.6 times the best possible instead of 1.3.
• Exponentials: forecast the logarithm. Every model's error grows with the growth rate. On log x, Timestral stays between 1.1 and 2.1 times the best possible.
• Trust the band within 1 024 steps only. Beyond one decoded block, the median is fed back and the band collapses. TimesFM 2.5, past 1 024 steps, does the same from step 129.
• Shift a leading series by its lead. As is, it is ignored. Passed as a known covariate, it cuts the error to 4 % when its lead covers the horizon.
• Level shifts and gaps are handled. With less than two periods of a slow cycle, TimesFM 2.5 does better, and a chirp defeats every model.
• Keep the mirror averaging. Without it, the error on sines rises by 5 to 29 %.
• For the model, both weak points sit in the decoding, not in the network: the default clip, and the median fed back between blocks. Rolling out from quantile paths, as Chronos-2 does, would likely keep the band.