TimesFM 2.5
A time-series forecasting model released by Google Research that uses a two-stage approach to handle covariates and generates forecasts block by block.
Timeline (last 90 days)
In a what-if scenario where the marketing budget was doubled, TimesFM 3.0 showed almost no reaction while 2.5 predicted a large jump in revenue; the agent trusted 3.0 and recommended Option A, but the real outcome was that Option B brought in more revenue, and the agent chose wrong in two of three test rounds.
TimesFM 2.5 uses a two-stage approach to handle covariates: first fitting the covariate effects with a regression outside the model, then letting the model predict the residuals.
In a test on daily New York subway ridership, TimesFM 2.5 and 3.0 reduced forecast error by about 25% relative to a seasonal baseline under zero-shot conditions.
On the New York subway data, TimesFM 2.5 made slightly better use of additional future inputs (such as a holiday calendar) than 3.0 and had more reliable uncertainty intervals, while 3.0's inference speed was several times faster.
In the Lumen business simulation, TimesFM 3.0 completed a revenue forecast for the next two months in under a second with a deviation within 2%; when re-evaluated a quarter later, the deviation was still within 6%, and neither version exceeded its respective uncertainty interval.
TimesFM 2.5 has about 200 million parameters, 20 transformer layers, a model width of 1280 with 16 attention heads, an input patch of 32 steps, a single output of 128 steps, and a maximum context of 16384 data points.
The TimesFM 2.5 weights are hosted in safetensors format on Hugging Face and are 925 MB in size.
The community project webTimesFM runs TimesFM 2.5 entirely in the browser, with no server, Python, or registration required, and data never leaves the local machine.
When TimesFM 2.5 was released, both Flax (JAX) and PyTorch implementations were provided.
TimesFM 2.5 removed the frequency setting and supports covariates via XReg.
All Facts (10)
In a what-if scenario where the marketing budget was doubled, TimesFM 3.0 showed almost no reaction while 2.5 predicted a large jump in revenue; the agent trusted 3.0 and recommended Option A, but the real outcome was that Option B brought in more revenue, and the agent chose wrong in two of three test rounds.
50%UnverifiedTimesFM 2.5 uses a two-stage approach to handle covariates: first fitting the covariate effects with a regression outside the model, then letting the model predict the residuals.
50%UnverifiedIn a test on daily New York subway ridership, TimesFM 2.5 and 3.0 reduced forecast error by about 25% relative to a seasonal baseline under zero-shot conditions.
50%UnverifiedOn the New York subway data, TimesFM 2.5 made slightly better use of additional future inputs (such as a holiday calendar) than 3.0 and had more reliable uncertainty intervals, while 3.0's inference speed was several times faster.
50%UnverifiedIn the Lumen business simulation, TimesFM 3.0 completed a revenue forecast for the next two months in under a second with a deviation within 2%; when re-evaluated a quarter later, the deviation was still within 6%, and neither version exceeded its respective uncertainty interval.
50%UnverifiedTimesFM 2.5 has about 200 million parameters, 20 transformer layers, a model width of 1280 with 16 attention heads, an input patch of 32 steps, a single output of 128 steps, and a maximum context of 16384 data points.
50%UnverifiedThe TimesFM 2.5 weights are hosted in safetensors format on Hugging Face and are 925 MB in size.
50%UnverifiedThe community project webTimesFM runs TimesFM 2.5 entirely in the browser, with no server, Python, or registration required, and data never leaves the local machine.
50%UnverifiedWhen TimesFM 2.5 was released, both Flax (JAX) and PyTorch implementations were provided.
50%UnverifiedTimesFM 2.5 removed the frequency setting and supports covariates via XReg.
50%