Software Engineering UKSoftware Engineering UK
Predicting the Future with Zero-Shot Models
Software Engineering UK

Predicting the Future with Zero-Shot Models

Alexander HillWords by Alexander Hill

Updated

Further up the stream

Dropped onto a retailer's sales exports without any training, a pretrained forecaster lands close to the model an analyst spent a season tuning. The same forecaster, shown the price change the board is weighing, can return a line that barely fluctuates. Both results come from one model on one afternoon, and the reason they do not contradict each other is that they answer different questions.

Google Research's release of TimesFM-3 is the occasion for taking that pair seriously. The family has been public since 2024, and this is the first version built for multivariate forecasting from the start. Google Research says the model is pretrained for the multivariate case (Google Research). That distinction is where the useful engineering resides, and it is also where the disappointment usually comes from.

What freezing the parameters actually buys

Zero-shot forecasting means the weights are fixed at the moment the forecast is made. EmergentMind's survey of the capability states the constraint without softening it: no gradient-based fine-tuning, no dataset-specific retraining, so the model must generalise to a domain that may differ in granularity, seasonality or statistical behaviour from anything it saw during pretraining (EmergentMind).

The exchange is worth naming plainly. You give up the ability to inspect a fit. You get a forecaster that works on the first day, on a series with too little history to fit anything sensible, and on hundreds of series through one interface. What replaces the fit is per-series normalisation, which strips scale out of the input so a model trained across many magnitudes can read a new series without being told what units it is in.

That is the honest version of the benefit: maintenance replaced by reuse. The model has not learned your business. It has seen enough shapes that it does not need to be taught what a weekly cycle looks like.

Why one series at a time is the incorrect format

Google's own illustration is ice cream sales for a retail chain, and it is the right one because it is so ordinary. Past sales alone do not explain a spike. Sales of cones and syrups, historical foot traffic, the weather forecast and a promotion calendar all bear on the number, and a model that uses only the target's own history has no channel through which any of them can arrive. Its forecast is a projection of habit.

This is the part of the announcement that matters to anyone running production systems, more than any benchmark position. A forecast that cannot move when the plan changes is not a decision aid. It is a rear-view mirror with a calendar attached. The multivariate turn is worth tracking because it decides whether the forecast can respond to something the business is about to do.

The three inputs a forecast can carry

Multivariate support is easy to assert and hard to make real, so it helps to separate the three things a model can be handed.

Target series. These are the numbers you want forecast. A model built for multiple targets predicts several of them jointly, so a promotion on one brand is visible to the forecast for another. Google Research says the model returns both point and quantile forecasts for every target.

Past covariates. These are series known only up to the present, such as historical foot traffic or last month's staffing. They cannot tell the model what is coming, but they help it separate an unusual past from an unusual pattern, which keeps the baseline it projects forward from being inaccurate.

Past and future covariates. These are known into the horizon itself: a promotion calendar, a holiday list, a weather forecast, a scheduled price change. They are the only input that can pull a forecast toward an event that has not happened yet.

The distinction between the second and third is the load-bearing one. A past covariate explains; a future covariate changes the answer. Most teams that try zero-shot forecasting and come away unimpressed have supplied the first and expected the behaviour of the second.

How the model keeps time and series apart

Google Research's description has the model cutting each series into patches of 32 time steps and emitting nine quantiles at each step of the horizon (Google Research). Patches make the sequence short enough to attend over, and quantiles make the output a distribution.

The architectural idea worth understanding is that two relationships are being learned at once and they have different rules about leakage. Within a single series, attention has to be strictly causal: a token may only see earlier tokens in its own series, or the model learns from the future it is meant to predict. Across series, at a given time step, attention should be unrestricted, because the current values of related series are legitimate inputs and the correlations between them are the whole point. Google Research alternates temporal and variate attention across the layers to keep the two rules separate.

The decoding step follows from that. The horizon is filled with masked placeholders, targets and past covariates are hidden behind them, future covariates stay visible, and the model completes the whole horizon in what Google Research calls a "single forward pass". Earlier versions generated a patch at a time, which compounds error and costs latency. Filling the horizon at once removes the loop, and keeping future covariates visible throughout is what lets a known event reach every step it affects.

Zero-shot performance has a maximum level

MindStudio tested the model on New York transit data and on a simulated software business with several years of daily history. On a two-month revenue forecast with no training, it landed within 2 per cent of the true figure, and a rerun a quarter later stayed within 6 per cent, with the outcomes inside the predicted uncertainty ranges (MindStudio).

Then the same test asked a planning question: double the January marketing budget. The newer model barely changed its forecast while the older one predicted a large jump, and an agent comparing the two proposals favoured the newer model and recommended the incorrect plan more often than not. The underreaction held across four different ways of encoding the spend variable, which rules out a formatting quirk. The constraint is structural. A zero-shot model can only extrapolate along a lever that has actually moved in the history you showed it. If the price has not changed in four years, no model, pretrained or fitted, can read elasticity out of a flat line.

Domain matters as much as lever. A global study of mortality forecasting across many countries and age groups found TimesFM consistently underperformed traditional methods, while CHRONOS was competitive at short horizons and improved with fine-tuning, and a Random Forest trained on the mortality data did best overall (Petnehazi et al.). A top rank among pretrained foundation models on public benchmarks is a statement about that population.

What has to be in front of the model

The COSMIC paper from AWS AI Labs and academic collaborators states the covariate problem exactly: Existing pretrained models either do not support covariates or fail to incorporate them effectively, and since no model is trained on the task, the relationship between a covariate and a target has to be inferred from the context supplied at inference time (Auer et al.).

That gives a practical rule with real consequences for how you prepare data. The model learns the promotion-to-sales relationship by seeing promotions move sales somewhere in the context you hand it. If that history contains promotions but never a quiet week to compare them against, or a budget lever that never varied, the relationship is not there to be found. Feeding a future covariate is necessary and not sufficient. The past has to contain the experiment.

Four guardrails follow, and they are unglamorous. Extrapolate only levers that have moved historically. Run small controlled experiments for levers that have not. Sanity-check large proposed changes using the lift you have actually measured. Backtest the uncertainty bands on your own history before relying on them, because intervals that run narrow are worse than no intervals at all.

Licensing also determines the configuration of the deployment. TimesFM 2.5 is released under Apache 2.0 and can be self-hosted commercially, while the TimesFM 3.0 weights are non-commercial to self-host, with commercial use routed through Google Cloud and BigQuery, MindStudio reports. Google Research says the code is on GitHub and Hugging Face, with BigQuery integration following, and that the earlier model is already reachable through the AI.FORECAST command.

Forecasting as simulation rather than a single line

There is a second way to use a zero-shot model that the release notes do not mention and that changes what you ask it. Lacuna's note on zero-shot prediction through generative future simulation focuses on models that generate many possible futures and derive a risk by counting how often the event of interest appears across them (Lacuna).

The relevance to a probabilistic time-series model is direct. A quantile forecast already represents a fan of outcomes. Once the model accepts future covariates, the interesting question stops being what sales will be and becomes how often a proposed promotion clears its target under the conditions you can foresee. That is a better fit for planning, and it is a use the multivariate model supports without further training.

What to do with the release

Treat a zero-shot model as two things and not a third. It is a strong baseline you can stand up in an afternoon, which is valuable because most forecasting estates have no honest baseline at all. It is a continuation engine for patterns already present in the data, and on that job it competes with maintained per-series models while costing far less to keep. It is not an oracle for decisions that push the business outside anything it has done before.

The practical test is one question asked of every input: Did this lever shift in the supplied history, and did the outcome shift with it? Where the answer is yes, the model can use the relationship, and the future covariate does the work. Where the answer is no, no amount of pretraining substitutes for a small experiment. The two facts from the opening are not in tension once you ask that question of each forecast in turn.

Sources

Source: TimesFM-3: A zero-shot foundation model for multivariate forecasting

The usual questions

What is zero-shot forecasting?

It is a method where model weights are fixed during forecasting. It requires no gradient-based fine-tuning or dataset-specific retraining.

How do multivariate models differ from univariate ones?

Multivariate models process multiple targets and covariates jointly. This allows them to respond to future events like price changes.

What is the difference between past and future covariates?

Past covariates explain unusual patterns in historical data. Future covariates, such as promotion calendars, can change the forecast.

Filed so far