Overview and Information About our Forecasting Approach
Our forecasting uses time series analysis, the type modeling most appropriate for demand forecasting in the majority of applications. This allows our system to pick up seasonal demand patterns, identify and discount anomalies, and recognize growth trends, giving you more accurate demand forecasts, and therefore, better purchasing recommendations.
The best seasonal forecasting results will be had once an item has 25 months or more of actual sales history in the system. By using data manipulation techniques described below, you can reduce the minimum to 12 months of actual sales, but often times we caution against this unless the demand pattern is likely to be very similar to the existing history in the future.
While it may initially seem like a limitation to not be able to recognize seasonality naturally before 25 months, this approach has a very important benefit. We are able to automatically discount past sales anomalies like out of stock periods and one-off promotions in future forecasts, ensuring the most accurate results statistically as well as realistically.
Note Regarding Sales History and Seasonality in General:
Items must have at least one month of sales history to be forecast at all and three full months of sales history to be deemed reliable for short-term, non-seasonal forecasts. ForecastRx stores up to 4 years of Sales History, and similarly to any form of data analysis, the more information, the better. The knowledge below will help you optimize forecasts for items with and without extensive sales history.
Having covered the basic principles above, let’s take a look at specific cases you may encounter with your items and how to address them.
Items With Upcoming Promotions
If you expect to run a promotion, and it isn’t something that is an annual occurrence, you will definitely want to manually account for that promotion during your ordering process. The best way to do this without affecting future forecasts negatively is by increasing your Purchase Quantity manually in the PO creation section of the application. If you are using QuickBooks, creating a sales order for the additional quantity with a due date in the month you expect to run the promotion is a great way to have ForecastRx account for this automatically.
Out of Stock Periods
Items that have had multiple recent out of stock periods, or items that have had unintentional out of stock periods for multiple months over the course of multiple years, will likely have less accurate forecasts. The ForecastRx algorithms can discount most of these occurrences, but the best way to address them, if they are prolific, is by using Modify Item History in the “Advanced” section under Manage. Here, you can manipulate historical data to better reflect the true demand an item should have experienced had a stock-out not occurred. Note for Amazon clients: since Amazon does not store specific data regarding stock-outs going back more than 90 days, we don’t have access to data that will tell us if a zero-sales month was due to an out-of-stock condition or just a lack of demand.
Data Manipulation Techniques in ForecastRx
Modify Item History
The Modify Item History in the Manage > Advanced section of the application offers three unique ways to modify your past sales data in order to improve your forecasting accuracy. Popular reasons for modifying sales history are: adjusting for stock-outs, adjusting for promotions, adjusting for other identified anomalies, forcing models to recognize seasonality, leveraging past sales history of a phase-out item for its replacement. Once your target item is in view, choose between these data modification techniques based on your needs:
Manual Overrides: Here, we give you a simple way to modify an item’s monthly sales data. We then use those adjusted values when forecasting your future demand instead of the quantity pulled from your data source. For example, if you went out of stock midway through a peak selling month, you can adjust the sales quantity for that month to more accurately reflect the true demand that item experienced. Some selling platforms, such as Amazon Seller Central, allow you to pull recent lost sales reports via CSV. These are a good place to start when trying to accurately correct Historic demand.
Copy History: Within the Modify Item History function we offer a Copy History technique used to copy one item's history over to another. This can be used for new product introductions or phase-outs where you may be replacing a version 1 item with a version 2 item. You can also use this if you believe a new child SKU introduction will have a similar demand pattern to an existing one.
Backfill: Within Modify Item History we also offer a Backfilling data manipulation that will help our models pick up seasonality for items that have less than two years of sales history. Be careful using this feature; it can help forecasts in some situations, but harm them in others, especially when there were uncorrected sales anomalies in the past.
In simple terms, you should look to modify items exhibiting historical stock-outs, drastic increases or decreases in demand, or items with consistent seasonal demand patterns that have less than two years of sales history. Each of these situations, if not addressed, can deprecate future forecast accuracy.