PlanIQ - Unexpected Forecast Drop After Removing Outlier Correction

seymatas1

I’m forecasting monthly demand with a focus on a rolling 6-month horizon (out of an 18-month view), using PlanIQ in Anaplan. My forecast uses 7 models: ARIMA, Prophet, ETS, MVLR, Ensemble, DeepAR, and CNN-QR.

Each month, I perform 6-month backtesting and calculate MAPE for each model, then select the best-fit forecast using 1- MAPE.

Until last month, I used an outlier correction method that automatically capped values using upper and lower bounds. However, this often cut off real business growth signals. So, I:

• Removed all outlier correction from the past 7 years of historical actuals
• Recreated the PlanIQ data collections and models using this new history

After doing this, I expected higher forecasts, since peaks are now kept.
But instead, I’m seeing:

• Lower forecasts
• In some cases, flat or even decreasing trends
• Even though the business is growing in certain regions and customers, and we’ve seen strong seasonal growth recently

Some ideas I considered:

• Models may be slow to recognize new patterns and will catch up after a few months
• Changing the historical data changed the statistical properties (mean, variance, skewness), which affected model behavior
• PlanIQ might have retained some internal state based on the previous (truncated) data, causing inconsistency

Additional context:

• No single model dominates; best fit varies across customers and regions
• I have a manual override process for model selection, but it’s time-consuming and hard to scale
• I don’t segment customers by behavior (e.g., growing vs. steady-state)
• We recreate data collections and forecast models each time we run the forecast

My questions:

1. When I change historical data and recreate the collections, does PlanIQ retrain models from scratch on the new data?
2. Is there a way to tune model flexibility (e.g., to prevent underfitting or excessive smoothing) in PlanIQ?
3. Why would removing outlier correction (i.e., restoring peaks) result in lower forecasts, even when using MAPE-weighted model selection?
4. Is the 1-MAPE formula causing this unexpected behaviour? Should I use 1/MAPE? I am interested in accuracy for big customers.

Would love to hear if anyone has experienced something similar in PlanIQ or has thoughts on how to better handle this situation.

taranjassi

Hi Seyma,

"After doing this, I expected higher forecasts, since peaks are now kept.
But instead, I’m seeing:

• Lower forecasts
• In some cases, flat or even decreasing trends
• Even though the business is growing in certain regions and customers, and we’ve seen strong seasonal growth recently"

I'm curious to know whether this happened for all forecasts, or only a select few of the ones you normally generate predictions with?

My thoughts are that due to including the outliers, increases and decreases are inflated, and therefore the trends and seasonality have fundamentally changed in the historical actuals. This could cause the weights associated with particular algorithms to change, or become more negative. As a result, PlanIQ could infer a more significant 'drop' due to the inclusion of outliers and generate a prediction which flat-lines or is lower than expected.

I know this can be the case in LSTM models and generally in forecasting, where outliers or 'demand spikes' can distort normalisation of values when preparing data for forecasts, resulting in 'noiser' than normal predictions.

Keen to hear thoughts from others on this and if anyone's had similar experiences!