TimesFM 3.0 Cuts University Building Energy Forecast Error by 43%

Building operators need day-ahead load forecasts to plan heating, ventilation and other services in advance. The authors of a new paper on arXiv used data from a university building to compare statistical models, classical machine learning, a recurrent network and the pretrained TimesFM model. TimesFM 3.0 with covariates had the lowest overall errors: its mean absolute error (MAE) was roughly 43% below that of the weekly seasonal naive forecast. But XGBoost did not fall far behind, and paired daily comparisons did not confirm a clear MAE advantage for TimesFM.

How the benchmark is set up

The authors compared the models on more than accuracy. The comparison also covered the input data required at forecast time, the need for local training and computational cost. For an operator these are not academic details: they determine how much engineering work a deployment will take.

Validation is chronological, meaning the model never sees the future during training. All configurations were evaluated on the same 57 test days, and the results were broken down by working days, weekends and holidays. Holidays are the hardest regime for such buildings, because the load differs sharply from normal.

There is one caveat about weather. The covariates came from retrospective observations, not from the actual forecasts an operator would have received the day before. Weather forecast errors did not enter the experiment.

What the numbers show

• TimesFM 3.0 with covariates: the lowest overall errors, with MAE roughly 43% below the weekly seasonal naive baseline.
• XGBoost: remains competitive. Paired daily comparisons show no clear MAE advantage for TimesFM.
• TimesFM on consumption history only: also a strong result, with no weather or other inputs.
• LSTM: takes longer to train and did not beat either TimesFM or XGBoost.
• Training: TimesFM requires no weight fine-tuning for the building, while XGBoost trains quickly and produces forecasts locally.

A separate note on covariates. For TimesFM, they improve the holiday forecast but slightly worsen it for weekdays and weekends. Overall MAE barely changes. The extra inputs work as a targeted tool for a rare regime, not as a universal improvement.

A trial historical evaluation showed seasonal swings in quality. Comparing weather data sources produced only small changes in the accuracy of the chosen TimesFM configuration. The choice of weather provider, at least here, is secondary.

How to choose a model for a building

The authors draw a cautious conclusion: the pretrained TimesFM is a strong starting baseline for this building, and XGBoost is a competitive alternative that is tuned locally. The result applies to a single building and 57 test days, so it should not be carried over to other sites without verification.

In practice, the approach looks like this. If you need a working forecast quickly, without a training cycle, start with TimesFM on consumption history only. If the team is ready to maintain a local model and values a fast training and inference cycle, XGBoost remains a sensible choice. In this comparison, LSTM gave no reason to pay for its longer fitting time.

Add covariates when holidays are critical for the building and preparing weather data is justified. If weekdays and weekends matter more, the gain from covariates is doubtful. Before deployment, repeat the test with real weather forecasts instead of retrospective observations. This is exactly where the study's results come closest to the limit of their applicability.