Series out of your own model
The input is a numeric attribute of your objects over time. No separate import, no parallel data set to keep in sync.
Forecast
Time series from your own object data, forecast with nine methods and backtested — so you know which one to trust.
The problem it solves
Forecasting tools usually offer one method and present its output as the answer. On a short or noisy series, though, a naive baseline can beat a sophisticated model — and you have no way of knowing which happened without checking.
- Your data — any numeric attribute over time, straight from your model
- Nine methods — from a cheap baseline to Prophet
- Backtesting — measured accuracy per run, not a promise
Plainova runs the methods against your actual history, scores them, and shows you the score. Picking a method becomes a decision you can defend.
What it does
Backtesting and a quality rating
Every run is scored against held-back history and rated — not evaluated, poor, fair, good or excellent — so a method has to earn your trust on your data rather than in a brochure.
Confidence bounds
Where the method supplies them, the forecast carries upper and lower bounds instead of one deceptively precise line.
Runs are kept
Every run is stored with its points, method and score. You can go back and see what you predicted — and what actually happened.
Calculated in the background
A run is queued and processed by a worker, and you are notified when it completes. A long series does not block the interface.
Prophet with a fallback
Prophet runs in a separate Python service. If that service is unavailable the run falls back to Holt-Winters instead of failing.
The nine methods
- Naive
- last observed value carried forward — the cheap baseline
- Seasonal naive
- last value of the same season, for weekly or yearly patterns
- Moving average
- average of the last N observations, smooths noise on flat series
- Simple exponential smoothing
- weighted average that reacts faster to recent levels
- Holt linear
- level plus trend, projects a straight trend line
- Holt-Winters
- triple exponential smoothing: level, trend and seasonality
- Linear regression
- least-squares straight-line trend with confidence bounds
- ARIMA
- captures autocorrelation and trend through differencing
- Prophet
- automatic trend change-points and multiple seasonalities
What a forecast cannot tell you
Stated plainly, because a forecast presented without its limits is a number people act on.
It extends the past
A change nothing in the history hints at — a new product, a lost customer, a price move — will not appear in the projection. No method fixes that, and one that claims to is fitting noise.
Short histories flatten everything
With too few periods every method degenerates towards the average. Seasonal models start earning their keep at roughly two full cycles; below that they mostly add confidence, not accuracy.
Data quality beats method choice
Gaps and outliers move the result more than picking Holt-Winters over exponential smoothing does. Fixing the series is usually the better hour spent.
The quality figures are a comparison, not a guarantee
MAE, RMSE and MAPE are measured against history the method was not shown, which makes them honest — and still a statement about the past.
Where the AI comes in — and where it deliberately does not
The forecasting itself is statistics, not a language model, and that is on purpose: these methods are reproducible and their accuracy is measurable. A language model that simply asserts a number is neither.
The AI sits around it instead. You can ask for the series you want in natural language, and forecast results are ordinary data in your model — so the same search, workflows and API apply to them as to everything else.
Which numbers would you like to see forecast?
Bring a series with real history. Whether it forecasts well is something the backtest answers in the first session.