Google's GPT-3 Moment for Time Series & Forecasting (TimesFM 3.0)

Author: Prompt Engineering · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Google’s GPT-3 Moment for Time Series & Forecasting (TimesFM 3.0) · Channel: Prompt Engineering · 13:42 · Watch on YouTube

Topics: Enterprise AI applications, AI evaluation and benchmarks, Data, databases and embeddings, AI models

Summary

Google’s TimesFM 3.0 is a time-series foundation model designed to produce zero-shot forecasts, including forecasts informed by future inputs such as holidays and promotions. In tests on New York transit data, TimesFM 2.5 and 3.0 both beat a seasonal baseline, while 3.0 ran faster. In a simulated business, the models forecast existing revenue patterns well but disagreed sharply about the effects of a marketing budget change, leading an AI agent to choose the wrong plan in two out of three runs. The video concludes that these models can help forecast established patterns, but what-if planning requires caution, historical variation, backtesting, and attention to licensing.

Key points

TimesFM is a decoder-only transformer trained on time-series data that predicts future chart patches from historical patterns.
Google says TimesFM 3.0 was pretrained on more than a trillion time points across real and synthetic datasets.
TimesFM 3.0 can incorporate future inputs and multiple related series, such as planned holidays or promotions.
On New York transit data, both versions reduced forecast error by about 25% compared with a seasonal baseline; adding holiday information improved them further.
In a simulated business test, TimesFM 3.0 forecast revenue within 2% for a two-month period and within 6% on a later forecast.
The two versions gave opposing predictions about the effect of doubling January marketing spend, and an agent chose the wrong plan in two out of three runs.
A model cannot reliably predict the effect of a business lever, such as price, if that lever has never varied in the available historical data.
Large changes beyond historical experience can make the model versions diverge; compare forecasts with past campaign results and backtest uncertainty ranges.
According to the video, self-hosted TimesFM 3.0 weights are non-commercial, while commercial access is available through Google Cloud and BigQuery; TimesFM 2.5 is under Apache 2.0.

Chapters

Start	Chapter	Summary
00:00	What TimesFM is and how it works	The video introduces zero-shot time-series forecasting and explains TimesFM’s transformer architecture, patch-based predictions, and training scale.
03:40	What changed in TimesFM 3.0	The video compares version 2.5’s separate handling of future inputs with version 3.0’s ability to use them inside the model, then tests both on New York transit data.
06:51	Testing business forecasts	A simulated software business and isolated database branches provide a setup for testing revenue forecasts and what-if plans.
08:57	When what-if forecasts disagree	The models disagree about marketing spend, and the agent’s selected plan underperforms the alternative in the simulation.
11:05	Limitations and conclusions	The video outlines four production cautions—unobserved levers, out-of-distribution changes, uncertainty calibration, and licensing—and concludes that these models are more dependable for existing patterns than novel planning decisions.

Quotes

The easiest way to think about it is an autocomplete for charts. — 02:35

When you predict a plain forecast with history alone, the model predicts a normal Monday on Labor Day — 05:15

The agent trusted 3.0 and recommended proposal A. — 10:35

The model can only extrapolate levers that have actually moved in your history. — 11:38

For continuing your existing patterns, the answer is yes. — 13:12

Concepts and entities

TimesFM, TimesFM 2.5, TimesFM 3.0, Google, Neon, Gemini, ARIMA, zero-shot forecasting, time series forecasting, covariates, Neon branches

Full transcript (original language)

00:00 Can you predict future with AI? Time series forecasting models have been doing that. This includes predicting things like daily revenue, cloud server loads, or inventory demand. And it’s one of the biggest real-world application of traditional machine learning. While foundation models have completely taken over language, code, images, and video, most forecasting teams still have to train and maintain a separate classical model for every single series they track. Google recently released Times FM 3.0 to change that. It is a foundation

00:31 model built specifically for time series that forecasts zero-shot without any training on your own data. And the third version uh can now look ahead at things you already know about the future, like an upcoming holiday or a planned promotion for a business. So, forecasting has real-world applications. Here it is running on a real daily ridership from New York Subway, where forecasting the next 4 weeks in a fraction of a second. Zero-shot without any training. Now, the big question is

01:02 whether a model that has never seen your data is actually reliable enough to run a business on. So, I tested two versions of this model. And turns out they can beat classic baseline out of the box and landed within 2% on a revenue basket. The two versions give me completely opposite answers. Which is a very important caveat if you want to use this model in production. So, in this video, I want to show you how Times FM work under the hood, what

01:32 change in this new version, and where a zero-shot forecast breaks when you use it to make planning decisions. So, let’s start with what exactly Times FM is and what it’s not. So, it is a decoder-only transformer from Google research trained exclusively on time series data. Now, compare that with how forecasting works in most companies today. You can take one series, fit a model like ARIMA or a gradient boosting tree, then you train tune it, and then do that all over

02:04 again for the next series. So, theoretically, if you have 1,000 stores, this is going to mean that you need to train 1,000 separate models to maintain this. Times FM is trying to change it. Now, the obvious objection is that one general model cannot possibly know your business better than a model fitted to your own history. And that is actually a fair point, except a forecasting model does not needs to know your business. It needs to recognize the shape of your numbers or patterns. It needs to be able

02:35 to look at weekly cycles, holiday dips, slow trends, and sudden spikes. And these are the things that repeats across almost every domain. To read those patterns and shapes, Times FM cut your history into patches of 32 time steps, and each patch becomes a single token. The easiest way to think about it is an autocomplete for charts. Just like a language model predicts the next tokens in a sequence, Times FM predicts the next patch of your chart. And instead of returning a single line,

03:07 it outputs a median forecast plus an uncertainty range around every single day. So, you can see both the likely case and how bad this downside could get. The reason it can do this zero shot is scale. According to Google, a Times FM 3.0 was pre-trained on more than a trillion time points across both real and synthetic data sets. So, by the time it sees your series, it has already learned how those cycles and spikes behave in this humongous data set. Let’s

03:40 look at what actually changed in this new version. Uh because I’m going to be running both the 2.5 and 3.0 side by side, and they handle future plans in two two different ways. Times FM 2.5 came out last year, and by itself, it only forecasts one series at a time from past values. If you want to account for a future holiday or a promotion, you have to run a separate step outside the model that fits a simple regression on those inputs first, and then Times FM

04:10 forecasts whatever is usual the regression did not explain. In version 3.0, that entire mechanism moves inside the model. You can feed it multiple related series at once along with past covariates and future inputs you already know ahead of time. Inside the transformer, each token looks ahead at those upcoming inputs. So, the model sees a holiday or a promotion coming before it predicts that week. And it generates the entire forecast horizon in a single pass instead of chunk by chunk.

04:43 Now, on paper, Google says 3.0 ranks first among foundation models across three public forecasting benchmarks. So, let’s test that claim on a real-world data that could not possibly be in its training data. This is the daily ridership for New York subway, buses, and commuter rails. I’m going to be testing both of these models side by side because there are some really interesting things about not only how the models behave, but also their licenses. So, I give both versions the history up to the end of August and ask

05:15 for the next 4 weeks. Now, we picked that cut-off on purpose. Times FM 3.0 was trained at the end of August, which means September couldn’t exist in the training data, and the specific date to watch here is the Labor Day. When you predict a plain forecast with history alone, uh the model predicts a normal Monday on Labor Day because nothing in the recent weeks tell it it that the holiday is coming. When I pass in the holiday calendar as a future input, the

05:47 forecast immediately drops for that Monday. Now, if you look at it, it doesn’t drop all the way to the bottom, but the model clearly recognizes that Labor Day is different. So, you can see contextual information is critical for forecasting. Here’s another experiment. I ran this across six cutoff dates and all seven transit services. Now, without any trading data at all, uh both versions cut the forecasting error by about 25% compared to a seasonal baseline. And

06:20 adding the holiday calendar improved both models further. On this data set, the two versions were surprisingly close. Version 2.5 actually made slightly better use of the extra inputs and had had more reliable uncertainty range. While version 3.0 ran several times faster. So, for continuing existing patterns in real data, you can get a strong forecast out of the box. And if you want more empirical tests on the foundation models for structured data and time series like this, make

06:51 sure to subscribe to the channel. Because a lot more on time series is coming. Now, a forecast only becomes useful when it runs on your own business data every day. And when you can test what if plans without touching production. This is where today’s sponsor, Neon, comes in. Neon is the backend platform for apps and agents, which combines Postgres with auth, storage, and scheduling functions in one place. Okay, so we’re going to look at a test bed that I created. Now, all of the

07:24 company’s history and future plans live in Postgres on Neon. And whenever an agent wants to test a what if scenario, it spins off an isolated Neon branch, a full copy-on-write clone of the database in about a second. Uh you can try it out using the link the description. More on that later in the video. To see what happens when an AI agent uses Times FM to make real business decisions, I simulated a multi-region software business called Lumen with

07:55 almost 4 years of daily trials and revenue, plus a planning table that tracks marketing spends, promotions, list prices, and holidays. Now, because I built this simulation, I also know the exact ground truth for what happens under any plan, so we can grade the models fairly here. First, I wanted to verify a baseline accuracy with zero data leakage. So, I branched the database as it looked on August 31st, where September October

08:26 do not exist. And we ran both models straight from Postgres on Neon. So, Times FM 3.0 forecasts the next 2 months in under a second and it landed within 2% of the actual revenue, which is pretty great. And then, I set up a nightly pipeline where a scheduled Neon function queues the job. And when I created the October 31st forecast a full quarter later, Times FM 3.0 was still within 6% on revenue, with both versions staying

08:57 inside their predicted ranges. Because the baseline forecasts uh looks so accurate, it’s very tempting to trust the model on strategy questions. And that brings us to the planning question from the start of the video. Now, in our scenario, marketing has two proposals for the upcoming quarter, but you can only fund one. So, proposal A makes the Black Friday promotion deeper and uh suggests or recommends to make it twice as long. Proposal B keeps Black Friday as is and

09:30 doubles the January marketing budget instead. At the same time, finance wants to know what happens if you raise our product’s price on January 1st. These are kind of the business decisions that you need to make and create models to forecast what is going to happen. So, in our case, I handed this exact decision to Gemini agent. For each proposal, the agent creates a separate Neon branch, updates the future plan table on that branch, and runs both Times FM 2.5 and

10:03 3.0. So, production data is never we are able to run the experiments on separate branches. This is a great feature that Neon offers. Okay, so let’s look at what exactly happened. On the price increase question, the agent actually did the right thing. It inspected the historical table, saw that the list price had never changed once in 4 years, and told finance not to trust any forecast for that price hike. But, on January marketing budget, the two versions of

10:35 the Times FM gave completely opposite answers. When you double the January spend, Times FM 3.0 barely reacted at all, while version 2.5 predicted a big revenue jump. The agent trusted 3.0 and recommended proposal A. When I checked the ground truth, proposal B actually generated more revenue, and the agent picked the wrong plan in two out of three runs. Now, the reason the agent got tricked is that the true outcomes sat right in the

11:05 middle. Version 3.0 underestimated the budget jump, while 2.5 overestimated it. And to make sure this was not just a quirk of how I formatted numbers, I tested four different ways of encoding the marketing spend, and Times FM 3.0 under reacted every single time. That brings us to the four limitations you need to know before you put time series foundation models into production because this is where a confident forecasting can cost you actual money.

11:38 First, the model can only extrapolate levers that have actually moved in your history. If a variable like price has never changed in your past data, no zero-shot model can know how your customers price elasticity. So, you’ll need to run small controlled experiments instead. Now, to be fair, same can be said about traditional machine learning models because they are fitting to your data. Second, watch out for large out of distribution moves. When a proposal pushes a budget or promotion much

12:09 further than your historical range, version 2.5 and 3.0 can diverge sharply. So, always sanity check a what-if forecast against your past campaigns lift rather than letting a single model version pick the winner. Third, calibrate the uncertainty range. On the New York Subway data, version 3.0’s confidence intervals were slightly too narrow. So, always run a backtest on historical branch first to check how often reality lands inside the

12:41 band. And fourth is the license. The weights for A Times FM 3.0 were non-commercial if you’re self-hosting them. Now, you have commercial access routed through Google Cloud and BigQuery. Whereas version 2.5 is open under the Apache 2.0 license. Okay, so coming back to our core question. Can foundation models that has never seen your data forecast it well enough to trust it? For continuing your existing patterns, the answer is yes. Both versions beat classic baseline zero-shot

13:12 and stayed inside their ranges a full quarter out with 3.0 being significantly faster. But, for what-if planning decisions, a time series model is only as good as the historical variations that are available inside your data. So, if you’re working with forecasting or structured data in production, let me know how you handle what-if scenarios in the comments below. And I’m putting together more of these tests on business data. So, if you want to see the next

13:42 one, make sure to subscribe to the channel. Now, time series is only half of the structured business data. In a recently tested how foundation models handle tabular data in this video right here on the screen. Anyways, I hope you found this video useful. Thanks for watching. And as always, I’ll see you in the next one.