Summary
The StockIQ forecasting algorithm is an in-house forecast algorithm developed by StockIQ, using our collective 50+ years of experience forecasting time series. It is actually a combination of several different forecasting approaches, something known as an "ensemble" algorithm, because it is a combination of several forecasting approaches. As a result, it is largely self-managing. However, you can tweak parameters when necessary to adjust its behavior. It has specific modes for:
- Non-Seasonal Items, or items with insufficient history for seasonality
- Highly Seasonal Items
- Brand new items
- Highly Volatile/Trending items
- Short-Life-Cycle Items (see Life Cycle features)
Further, it will try to borrow information from parent nodes within your forecasting hierarchy to help extract seasonality trends based on parent/grandparent/great-grandparent seasonality behaviors.
Settings
Below are some of the settings that you can modify to tweak the behavior of the algorithm and its component parts:
Volatility
The volatility setting determines the length of time that StockIQ uses when creating its trend line. When set to "Auto" StockIQ will always try to use the longest period of time it can - usually a 6 or 6 and 12-blended month average - to create the smoothest forecast possible, this corresponds to the "Very Low" and "Low" volatility settings When less data is available, StockIQ adjusts to a shorter horizon automatically, so that the volatility is sensible for new parts.
Setting the volatility to higher settings results in more reactive (aka volatile!) forecasts. historical forecast error will usually go down, but the additional reactivity can negatively affect buying - use this cautiously.
Very Low Volatility Option
Limited Auto Volatility: This enables an extra-stable version of the StockIQ algorithm, that looks at no less than 12 months of history for its smoothed averages during the trending step. Enable this if StockIQ is following changes more closely than you want. This can be useful for items with very mild seasonality that you don't want StockIQ to project a continued downward trend when the item is out of its season.
Bias Detection Option
Limited to Auto Volatility: Bias detection works by looking at recent data points, and seeing if the recent statistical model is consistently above or below the actuals by more than 5%. If the forecast shows bias in 75% of the last 12 months, 6 months and 3months, volatility will be raised to Low, Medium, and High, respectively.
Damping
This setting determines how quickly the trend "tails off" as you move out in time, on the assumption that no trend continues forever. It is rare to need to change this.
Fast Start Enabled
The "Fast Start" mode of the StockIQ algorithm allows StockIQ to take a short/instant look at a brand new part. For example, when you first sell an item, say 100 in a month, StockIQ will assume that this is the first sales period, and will begin generating forecast from this single point, rather than assuming there was a "smooth" of previous zero months. In this way, the initial forecast for an item will ramp up quickly based on its initial sales, without having to develop a trend history. It is recommended to leave this option enabled.
Applies To
Based on settings in the Seasonality Settings screen StockIQ Algorithm will figure out if a seasonal model and parameters should be applied.
If the StockIQ algorithm determines seasonal models should be used, the standard 3-sigma filtering is dropped to allow seasonal peaks and valleys to apply, since it is common for a highly seasonal item to have a very high period-to-period variance from when it is in versus out of season. Since the year-over-year trend line is calculated first to determine the direction of the part’s overall demand, it can be useful to tune how many years of history to look at for determining that historical trend direction to make sure the trend is not too optimistically up or pessimistically down. Once the trend line is calculated using Linear Regression, the seasonal, monthly coefficients are applied to the trend line.