Next Tuesday could resemble last Tuesday. This plain assumption is a serious opponent for any impressive forecast curve. If an AI is to estimate what lies ahead for a course room, a series of measurements or the demand for materials, it has to do more than extend the existing data with a smooth line.
Mistral introduced Timestral on 15 October 2026 as a family of time series models. The announcement describes probabilistic forecasts and the use of additional variables, both past ones and future ones that are already known. Vendor benchmarks are no substitute for a comparison on a series of your own that you understand.
The order is part of the information
A time series contains values that are tied to points in time. For a forecast, it is not only the numbers that count but also the intervals, the order and what was known at the time. An event that only became known later must not slip unnoticed into the earlier forecast.
Future information that is already known could be, for example, a public holiday that has already been published. A disruption identified after the fact does not belong in the same category. Anyone who mixes the two lets the forecast work with knowledge that would not yet have been available in real use.
An artificial series and a fixed cut-off
To get started, you can use an entirely invented daily series covering four weeks. The first three weeks form the available past. You hold back the last week for the comparison. The numbers here serve as teaching data and do not describe any real operation.
wochen = [
[10, 18, 11, 12, 16, 5, 4],
[11, 19, 12, 12, 17, 6, 4],
[12, 20, 12, 13, 18, 5, 5],
[12, 21, 13, 14, 18, 6, 5],
]
vergangenheit = sum(wochen[:3], [])
ist = wochen[3]
vorwoche = vergangenheit[-7:]
mae = sum(abs(a-b) for a, b in zip(vorwoche, ist)) / len(ist)
print("Mean absolute error of the previous-week rule:", mae)The reference rule takes, for each day, the corresponding value from the last known week. The example can be run with Python without access to a model. It produces only a comparison value for this artificial data; it does not measure Timestral.
Only a model forecast that has actually been generated would stand alongside it as a second estimate. It may use only the past that has been released for it and additional information known at that point. Record the cut-off explicitly so that nobody accidentally passes the held-back week in as input.
Several time windows instead of one lucky week
A single held-back section can be particularly easy or particularly unusual. For a later, more meaningful comparison, you shift the cut-off later, step by step. Each forecast only ever sees the past as it stood at that time. This principle shows whether an advantage depends on a single favourable week.
Fix the metric. The mean absolute error treats deviations according to their size. For a specific decision, however, estimates that are too high and too low can have different consequences. Missing material and unused stock, for example, are different problems, even if the numerical deviation is the same.
A range is an additional statement
Probabilistic forecasts can describe several possible trajectories or ranges. A shaded area should therefore have an explained meaning. It is neither automatically a guarantee nor mere decoration around a centre line.
With output that is actually available, check which quantiles or probability details it contains. Label any chart accordingly. A wide range can be more useful for planning than a seemingly certain single figure if it honestly makes uncertainty visible.
What you can take away from the experiment
The first insight does not yet need a model ranking: a forecast must be judged against an understandable rule and without knowledge of the future. The small script provides an independent starting point for this. After that, you can check whether a more elaborate approach brings additional benefit for your period.
Timestral makes such specialised models an interesting piece of AI news. The practical question remains concrete: which forecast is needed, what information was available and which errors are relevant to the decision? Here, a cleanly drawn time boundary explains more than a particularly impressive curve.
Sources and status
Sources as of 25 October 2026. Introducing Timestral.
