Loading today’s stories
Build1 distinct publisher3 min readPublished
Google's new 330 million parameter forecaster takes covariates and related series zero-shot in one forward pass, which puts a lot of bespoke per-SKU plumbing up for review, assuming the accuracy transfers to your own series.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
The interesting part of a call is which columns you must be able to fill in for dates that have not happened yet. TimesFM-3 splits inputs into the target series, past covariates whose future values are unknown, and past-future covariates whose future values you claim to know [10]. In the horizon, target and past-covariate patches are masked while the past-future covariates stay visible, which is what carries a holiday flag or a scheduled promotion into the prediction [10]. A weather forecast passed in as a known future signal is still a forecast, with its own error, and the post lists weather forecasts alongside promotions and holidays as example future-known features [16].
The tokenisation is what puts the promotion on the right day. Points are grouped into patches of 32 steps and normalised per series to handle wildly different scales [6]. For past-future covariates, each token concatenates the current patch with future patches, a lookahead that lets the model peek at upcoming known signals [8]. Tokens pass an input residual block, then the stack alternates attention along time and across series for several layers [7]. The authors say the older patch-at-a-time decoding cost latency, compounding error and compute [9].
The promotion chart is Google's own illustration, and the mechanism behind the lift matters more than its size. The post says the relationship between promotion and sales lift is learned from the historical context, then applied to future days with planned promotions [13]. A SKU whose context window contains no promotions gives the model nothing to learn the lift from. New mechanics on new products remain your problem.
The accuracy claim is where I would spend the review. Google states that TimesFM-3 significantly outperforms other forecasting models across major benchmarks [14], and the text I have names no dataset, no baseline and no metric [15]. For that number to transfer, your covariates have to be genuinely known at forecast time rather than backfilled during evaluation, and your sampling frequency and series count have to sit somewhere inside a pretraining mix of more than a trillion real and synthetic points [2]. The comparison also has to be fair on inputs. A univariate baseline denied the promotion column is not evidence about your workload.
Output is nine quantiles per target series at every horizon step, the 10th through the 90th percentile [11]. That works out to 10 percentile points of spacing and nothing above p90 [1]. If your safety stock is set at a 97.5% service level, you are extrapolating a tail off a decile grid, which is a modelling decision you now own rather than one the model made for you.
What actually weakens is narrow. If a per-SKU pipeline exists to bolt a promotion dummy and a holiday calendar onto univariate models, the replacement is a column layout plus an evaluation harness, with no training run, because the multivariate mode is zero-shot [5]. Google went from strictly univariate at 2.5 to native multivariate in roughly eleven months [3][2]. Pipelines that exist for hierarchical reconciliation, intermittent demand or business overrides are untouched here, because the post does not discuss them [15]. In my context the right sequence is to keep the harness, run the zoo and the foundation model side by side on your own series, and delete only what loses.
Ranked by verification strength, evidence, and original report placement.
In Google's ice cream example, passing the planned promotion schedule as a past-future covariate yields a forecast that anticipates a roughly 20% sales bump on each promotion day, while the univariate model only projects the historical weekly pattern with no knowledge of planned promotions.
According to the post, the model learns the relationship between promotions and sales lift from the historical context, then applies that knowledge to future days with planned promotions.
Google states that TimesFM-3 significantly outperforms other forecasting models across major benchmarks.
TimesFM-3 was introduced in a research.google blog post dated August 31, 2026, credited to Ayush Jain and Rajat Sen, Research Scientists at Google Research.
TimesFM-3 has 330 million parameters and is pre-trained on a real-world and synthetic time-series corpus comprising more than 1 trillion time points.
Up to and including TimesFM-2.5, released in September 2025, the models were strictly limited to univariate forecasting, using only the history of a single time series.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 31, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
Decathlon swapped weekly DeepAR retraining for a six-month fine-tune of Chronos-21 distinct publisher
invest
Google says frontier models already know the facts they get wrong. That is a budget decision.1 distinct publisher
product
Swapping out the GPU leaves four more rack lines on Nvidia's invoice1 distinct publisher
build
Google's wearable biomarker agent is built to distrust its own predictions1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Detailed, but all of it from the builder
The mechanism is documented tightly enough to argue with — patch size, masking scheme, lookahead tokens, quantile head — and that specificity is why this does not score low. The accuracy story is thinner than it looks: three benchmarks named, average ranks plotted, no error value anywhere, and every figure traceable to the same Google post that is selling the model.
Nothing to count yet
Google gestures at uptake across retail, finance, observability, manufacturing, healthcare and the natural sciences and names not one user — and that sentence is about the family since 2024, not about this model. With no checkpoint location, no licence and no serving path in the post, there is no adoption to measure here, only an intent to be adopted.
Engineering checkable, accuracy asserted
'State-of-the-art' and 'significantly outperforming' are carrying more weight than the charts beneath them, and the memorable number — a 20% promotion bump — comes from Google's own ice cream illustration rather than a held-out retail set. The gap is moderate rather than wide because the structural claims are falsifiable on sight the moment weights ship: either the horizon comes out in one pass or it does not.
The vendor is the only witness
Google wrote the post, chose the benchmarks, picked the comparators, drew the plots and framed the conclusion that bespoke per-series forecasting pipelines are now optional. There is no counterparty in the room, and the commercial upside of that framing accrues to the same company that hosts the inference.
Sure what was built, unsure how well it forecasts
We can describe this model with some confidence: the technical account hangs together and is specific in the places vendors usually go vague. We cannot yet stand behind the performance claim, or say when and how anyone will be able to run it — one publisher, no replication, no numbers under the rank plots.