Limitations

Be aware of these constraints before putting timesfm-mcp outputs in front of stakeholders.

Baseline backend

Assumes stationarity after decomposition. The seasonal-naive + trend model works well on stable, regular series. It will miss trend breaks, structural changes, and irregular shocks (a product launch, a pandemic, a supply chain disruption).

Quantile bands are Gaussian. The σ√h band formula assumes normally distributed residuals. Heavy-tailed series (e.g. viral traffic spikes) will have underestimated tail risk.

Season detection has limited range. The autocorrelation scanner tests lags up to min(n/2, 60). If your series has a period longer than 60 steps (e.g. quarterly data over many years with 5-year cycles), it won't be detected.

Minimum data: 3 observations. Fewer than 3 points raises a ValueError. In practice, you need at least one full seasonal cycle for the seasonal component to be meaningful (e.g. ≥12 points for monthly data).

TimesFM backend

System requirements: ≥ 16 GB RAM · ~800 MB disk Not suitable for resource-constrained environments. If your machine has less than 16 GB RAM, use the statistical baseline — it requires no extra install and is production-ready.

Install: pip install "timesfm-mcp[timesfm]" — the first inference call downloads model weights (~800 MB from HuggingFace) and loads the model, which takes 30–90 seconds.

Not fine-tuned to your domain. TimesFM is a zero-shot foundation model. It's broadly accurate but not specialized. A well-tuned Prophet or ARIMA model, fit to your specific series with expert seasonality knowledge, may outperform it.

Context cap: 1,024 points. Longer histories are silently truncated to the most recent 1,024. For very long, high-frequency series, this may lose long-run patterns.

Quantile interpolation. TimesFM natively outputs decile quantiles (10%, 20%, …, 90%). Requested bands outside 80%/90%/95% are linearly interpolated, which is approximate.

General

No multivariate support. forecast takes a single series. Covariates (promotions, holidays, external drivers) are not supported.

No real-time data. The server is stateless; it doesn't fetch or cache historical data. You pass the numbers; it forecasts them.

Illustrative outputs. The model has no knowledge of your business. Treat its outputs as a starting point, not a decision. Always sanity-check against domain knowledge and use backtest to measure accuracy on your specific data before relying on the forecast.

No confidence in the uncertainty bands. A 90% band means "if the model's assumptions hold, 90% of outcomes should fall in this range." If your series violates those assumptions (heteroscedasticity, regime changes), the band is not well-calibrated.

What to do about these limitations

• Run backtest on your actual data before presenting forecasts to stakeholders. MAE and sMAPE on a held-out window tell you how accurate the model has been historically.
• Check context.volatility — high volatility means wide bands and lower confidence. Present ranges, not point estimates.
• Combine with domain knowledge — the forecast is an anchor, not a verdict. Apply known events (pricing changes, seasonal promotions) as adjustments on top of the model output.
• For critical decisions, treat the lower bound of the uncertainty band as the planning figure, not the point forecast.