IFS.ai Time Series Forecast Model

IFS.ai TSPT Level, Trend and Season is an AI-based time-series forecast model for Demand Planning. TSPT stands for Time Series Pretrained Transformer and is based on Google’s TimesFM forecast model. It uses historical time-series data for a forecast part to produce a system forecast that demand planners can review, compare, and adjust before finalizing forecast values.
In practice, this means that you can select a forecast part in the Forecast Workbench, apply the TSPT Level, Trend and Season forecast model, examine the resulting system forecast, and then set the final adjusted forecast according to your planning judgment.
The forecast quality depends on the availability and quality of historical demand data for the forecast part, the selected forecast horizon, the period granularity, and the planner’s review of the output before using it in downstream planning decisions.

Who This Feature Is For and Why It Matters

Enable Machine Learning in Demand Plan Server

This feature is intended for demand planners and other users who work with forecasting in Demand Planning and need an AI-assisted forecast option that does not require statistical modelling expertise.

What It Does

The model computes a forecast from a single historical time-series vector. The model card states that supported period lengths include weeks, months, and quarterly periods. The attached model card also identifies Demand Planning and Nexus Generic Forecaster as primary use cases, while the training material focuses on the Demand Planning workflow.

Capability Description
Input A single historical time series for a forecast part.
Output A resulting system forecast that can be reviewed and used to set the final adjusted forecast.
Forecasting approach Time-series pretrained transformer model based on Google TimesFM.
Typical user action Select a forecast part, set the forecast model to TSPT Level, Trend and Season, review the system forecast, and adjust the final forecast if needed.
Out of scope Multivariate forecasting is out of scope for the model card.

Where the Results Appear

After the TSPT Level, Trend and Season forecast model is selected for a forecast part, the resulting system forecast is displayed for review. The planner can then examine the forecast output and set the final adjusted forecast.

Before You Start

Before using TSPT Level, Trend and Season, make sure the basic setup and access prerequisites are in place.

How to Use TSPT Level, Trend and Season

General workflow

  1. Open the Forecast Workbench and select the forecast part that you want to forecast with TSPT Level, Trend and Season.
  2. Change the forecast model parameter for the selected forecast part to TSPT Level, Trend and Season.
  3. Examine the resulting system forecast.
  4. Set the final adjusted forecast according to the forecast result and business planning context.
  5. Finish the forecast review process.

Important model selection notes

  1. TSPT Level, Trend and Season cannot be used together with a season model in Demand Planning.
  2. Weekly profiles can be used to split the computed forecast into days.
  3. The model is designed for single time-series forecasting. Do not position it as a multivariate forecasting feature unless that is separately confirmed.

How It Works

TSPT uses a decoder-only transformer architecture for zero-shot time-series forecasting. The model treats time-series data similarly to language modelling by using patching, causal attention, and stacked transformer layers to capture temporal patterns and predict future patches.

Concept Cloud-friendly explanation
Pretrained model The model is already trained. It is used to produce forecasts from the historical time series supplied as input.
Patching The time series is broken into groups of adjacent time points so the model can process patterns over time.
Patch-based prediction The model predicts the future forecast horizon using future patches, rather than generating one individual time step at a time.
Model scale The attached model card states that the model has 200 million parameters.

 

Output Quality and Benchmarking

The source material describes benchmarking against existing Demand Planning forecasting approaches. The model card states that the M4 dataset was used for testing, with 500 monthly time series selected for a short-history test and 500 monthly time series selected for a long-history test. The most recent six periods of known history were removed and used as the target forecast values, while the earlier periods were used as model input.

Benchmark item Source-based draft wording
Accuracy comparison TSPT / Google TimesFM delivered an 8.5% improvement in forecast accuracy compared with the existing Demand Planning forecast model.
Win-rate comparison TSPT / Google TimesFM: 66%; Demand Planning Best Fit: 34%.
Demand Planning metrics MAPE, MSE

 

Limitations and Things to Know

Responsible Use

Use TSPT Level, Trend and Season forecasts as planning support, not as automatic final decisions. Demand planners should review the resulting system forecast, compare it with business context, and adjust the final forecast where appropriate. Product teams that deploy or expose the model should confirm that the use case remains within acceptable AI usage boundaries and that any personal data considerations are assessed at the use-case level.

Troubleshooting and Review Checklist

If you notice… Check…
No useful forecast result Confirm that the forecast part has enough historical data and that the Demand Plan Server is running.
Unexpected forecast shape Review input history, seasonality assumptions, and whether the final adjusted forecast should be manually corrected.
Performance concerns Consider that the attached model card identifies the model as more compute-heavy than existing Demand Planning forecast models.