Articles on: Inventory & Products

Inventory Forecasting

Vintaflow projects future demand for each product inventory using a statistical forecasting model. This article explains how to choose a model, tune it, and read the accuracy score.

Where to Find Forecast Settings

Navigate to Inventory and select a warehouse
Open the product inventory you want to configure
Click Forecast Settings

The dialog is titled Forecast Settings and configures the model and hyperparameters for that inventory record.

Configuration Mode

Choose how this inventory gets its model:

Mode	Behaviour
Preset Model	Pick an algorithm and configure hyperparameters yourself, or let the trainer auto-tune them
Inherit from Host	Mirror the forecast configuration of another product inventory in your account

Use Inherit from Host when several products behave the same way - configure one well, then point the others at it.

Available Algorithms

Algorithm	Best For
None	No algorithm applied. Bring your own forecast, or receive zeros by default
Intermittent	Sporadic, lumpy demand with frequent zero periods
Simple Moving Average	Stable, low-volatility products. Averages demand over a fixed lookback window
Holt-Winters	Exponential smoothing with additive trend and seasonality. Robust for stable seasonal products
Prophet	Meta's time-series model. Handles missing data, outliers, and multiple seasonality patterns automatically
SARIMA	Seasonal ARIMA. Adds seasonal differencing and lag terms for periodic patterns
STLF	STL decomposition followed by a Theta forecast on the seasonally adjusted series. Good for strongly seasonal products
Heuristic	Simple statistical baseline using mean, median, or last value

Choose Intermittent when more than 50% of your historical demand observations are zero. Continuous-demand models produce systematically biased forecasts on that kind of data.

Selecting an algorithm shows its description inline, so you can confirm the choice before saving.

Reading the Accuracy Score

When a model has already been trained, a banner at the top of the dialog shows the current algorithm and its SMAPE score (Symmetric Mean Absolute Percentage Error). Lower is better.

Score	Colour	Interpretation
0.000 - 0.500	Green	Good fit
0.501 - 1.000	Amber	Acceptable, worth reviewing
Above 1.000	Red	Poor fit - try a different algorithm

Selecting a new algorithm replaces the currently trained model and its score. The new model has to be trained before a fresh score appears.

Hyperparameters

Algorithms that accept hyperparameters show a Hyperparameters section once selected. Defaults are filled in for you:

Holt-Winters - additive trend, additive seasonality, 52 seasonal periods
Prophet - additive seasonality mode, yearly seasonality 10, changepoint prior scale 0.05, seasonality prior scale 10.0, changepoint range 0.8
SARIMA - p 1, d 1, q 1, no trend, seasonal order (1, 0, 1, 52)
STLF - Theta method, m 52
Heuristic - mean

The value 52 reflects weekly data over a yearly cycle. Change it only if your demand cycle is genuinely a different length.

Skip Trainer

Tick Skip Trainer to disable auto-hyperparameterization and performance index recalculation.

Leave Skip Trainer off unless you have deliberately hand-tuned your hyperparameters. With it off, Vintaflow tunes the model for you and keeps the accuracy score current.

Saving

Click Save to apply the configuration. The forecast is recalculated on the next run and the results appear in the inventory forecast chart and weekly table. See Inventory Management.

Tips

Start with the default algorithm and only change it if the SMAPE score is poor
Seasonal products need at least one full cycle of history before a seasonal model is meaningful
Products with sparse or brand-new history do better on Heuristic or Intermittent than on Prophet or SARIMA
Use Inherit from Host to keep families of similar products consistent

Updated on: 27/08/2026

Was this article helpful?

Thank you!