Can Google's TimesFM Predict Expenses on Single-Family Construction?
We tested raw TimesFM 2.5 against real construction expenses. The answer was no—but the best predictions show why time-series models are still worth exploring.
Can AI predict construction spend on real projects?
Google's TimesFM is an impressive time-series foundation model. It attempts to extrapolate the future from past observations by drawing on broad training across many kinds of time-series data.
We tested raw TimesFM 2.5 against actual expenses from ground-up single-family construction projects.
The short answer is no. Raw TimesFM 2.5 cannot reliably project expenses on individual construction projects.
The results are tantalizing, though. Some predictions were close. TimesFM 2.0 can be fine-tuned, and Google has published research on in-context fine-tuning for time-series foundation models. That approach works like few-shot prompting for a language model: provide relevant examples at inference time so the model can adapt without a separate, task-specific training pipeline.
Update, July 2026: Google's open TimesFM repository now includes a TimesFM 2.5 fine-tuning example using PEFT and LoRA. The zero-shot test results below are unchanged; fine-tuning is now a practical next experiment.
The experiment
We tested three forecast intervals against real project expense histories:
- daily predictions over a 30-day horizon;
- weekly predictions over a four-week horizon; and
- cumulative monthly predictions over a one-month horizon.
Daily forecasts performed worst. They underpredicted actual spend by 64% to 97%. Too many zero-spend days confused the model.
Weekly aggregation helped by smoothing some of those zeros, but the forecasts still underpredicted by 42% to 76%.
Monthly predictions were better, but not reliable enough to use for project planning. The average absolute error was 64%. Only 24% of predictions landed within 25% of actual spend, and the median prediction error was -23.5%. The best results generally came after the model had at least 12 months of project history.
Where does TimesFM 2.5 work?
TimesFM works best when a series has a consistent pattern, observations are well distributed, and the model receives a relatively long history.
We found that it could sometimes produce a good construction expense forecast. Realistically, that means raw TimesFM becomes potentially useful on a ground-up project only after about 12 months of data.
Even then, the project does not stand still. Construction phases change. The type, frequency, and amount of expenses change with them. When the project begins winding down, the historical pattern becomes a poor guide to what comes next.
The best and worst predictions
The contrast within a single project shows the problem.
For Project A, the best backtest came after month 12. TimesFM predicted $385,439 of cumulative spend against $352,496 actual—a +9.3% error. Its worst result came after month 8, when it predicted $235,740 against $445,489 actual, a -47.1% error.
Project B showed the same instability. After month 18, TimesFM predicted $297,696 against $311,714 actual, only -4.5% off. After month 10, it predicted $259,861 against $134,894 actual, a +92.6% error.
Those good predictions are what make the experiment interesting. The model can find a useful pattern. It just cannot find one consistently enough across a project's changing phases.
Construction expense data is lumpy
Construction expenses—especially at the scale of a single-family home—are lumpy.
Some days have no expenses. On another day, one payment can equal 5% of the entire project budget. A framing draw, window package, mechanical equipment delivery, or major material order can create a spike that does not resemble the days before or after it.
Raw TimesFM 2.5 is not good at forecasting that kind of sequence. Aggregating the data by week or month reduces the number of zeros, but it cannot remove the underlying phase changes and one-time events.
Does more history help?
Sometimes—but not enough.
The chart below plots prediction error against the months of project history provided to the model. A perfect prediction would sit on the 0% line.
More history produces some strong predictions, especially after 12 months, but the errors remain widely distributed. Large overpredictions and underpredictions continue even with 16 to 20 months of history.
This is the central result: more project history can help TimesFM, but history alone does not make raw zero-shot forecasts dependable.
What is Buildplus?
Buildplus is the payments, expenses and invoicing platform built for contractors running cost-plus jobs. Every payment, swipe and reimbursable expense stays tied to the project it belongs to.
What should we try next?
The next step is adaptation.
At the time of the original experiment, TimesFM 2.5 could not be easily fine-tuned. The practical options were to fine-tune TimesFM 2.0 or wait for the in-context fine-tuning work described by Google Research to become available in a production-ready model.
The current TimesFM 2.5 repository now includes a LoRA fine-tuning example, which gives us a direct path for the next test. Training or adapting the model against a larger set of real construction histories may help it recognize phase changes, sparse expense patterns, and the shape of projects as they wind down.
In-context fine-tuning remains especially promising. Instead of expecting a broad foundation model to understand residential construction from one project's numbers, we could provide examples from similar completed projects as context for the forecast.
Other construction series may be more predictable
Individual projects are inherently uneven. A construction business should be less so.
As a contractor manages multiple projects, some jobs wind down while others start. Those overlapping cycles can produce a steadier company-wide series than any individual project.
That makes several other forecasting targets worth testing:
- total weekly company expenses across all active projects;
- weekly labor spend across the business;
- weekly material purchases from company cards;
- longer-running projects with more observations; and
- project types that spend more consistently from month to month.
These series may give raw TimesFM enough regularity to produce useful forecasts without fine-tuning. They may also help a contractor anticipate company-level cash flow even when no single job is predictable.
The takeaway
Raw TimesFM 2.5 cannot reliably predict expenses on a ground-up single-family construction project.
Daily data is too sparse. Weekly data is better but still materially underpredicts. Monthly data occasionally produces excellent forecasts, but its errors are too widely distributed for contractors to depend on it.
The experiment is still encouraging. The best predictions show that there is signal in construction expense history. Fine-tuning, in-context examples, a larger real-world dataset, and more stable company-level series may turn that signal into something useful.
Stay tuned for more AI in construction research.