Start with a Stat Forecast. And Start Strong!
Set a Strong Baseline Forecast Up Front, and Revisit Often to Maintain Accuracy
By Ryan Rickard, Director SCMO2
If you spend a lot of time in your demand planning cycle manually entering and adjusting your forecast, then start strong! If you manually enter other inputs you receive for your planning, then start strong! If you use historical input or analysis to help guide you when making your forecast adjustments, then start strong!
If you start strong with a statistical baseline, you can minimize the manual manipulation and time-consuming tasks downstream. Past data is the best indicator of future performance. Using your historical demand (orders, shipments, deliveries, consumption, or usage) to forecast future activity is a great way to create your statistical baseline.
SAP tools like IBP Demand and APO Demand offer numerous statistical algorithms that can be applied to your historical demand to find a reliable a future forecast. Both allow you to also perform a ‘best fit’ modeling sequence which analyzes several different statistical algorithms to find the best result from a list of selected algorithms. If you do not already know the pattern or volume of the item being forecasted, this ‘best fit’ option is a good feature to help you identify the best algorithm to use.
In today’s post, I outline all the key aspects involved in identifying the proper statistical forecast for your demand planning, and emphasize how to ensure your planning remains accurate over time. If you put in the effort up front and ‘start strong’ you can maintain an optimal and effective planning effort indefinitely.
Time Buckets
If you have weekly and monthly data, you should consider both. There will be patterns in your weekly data that may not appear when aggregated monthly. For example, a customer that buys on the first week of every month may reflect in the demand pattern when aggregated weekly. To ensure you have enough product to fulfill that week-one demand, you need to properly plan for it and not spread that demand across all four weeks of a future month. Sometimes weekly data can have more variability than monthly data in which case it then makes sense to use the monthly demand patterns. We offer more detail on this topic in our IBP webinar titled ‘SESSION II: How Much is Enough?’
Forecast Level
When creating a statistical forecast, the levels of data that you are reading are important, such as product vs. product/customer vs. product/customer/location. It is important to generate your forecast at the level that is least variable and fits the operational historical pattern. Keep in mind that your goal is to find the best product/location forecast to ensure you have adequate supply to accommodate all customers from the right locations and distribution points. Sometimes analysis at too detailed of a level can cause patterns at each level that when aggregated do not make sense. We offer more on this topic in our IBP webinar titled ‘SESSION III: Super Model Forecasting.’
Frequency
Ideally you should generate a new statistical forecast frequently enough to incorporate the most recent sales and market activity into the demand stream. This allows you to react as quickly as possible to changing patterns. But how quickly can your supply chain, planning tool and coordinating business processes actually react to forecast changes? It is a common misconception that frequent (i.e. weekly) stat forecast updates will constantly result in big changes to demand signals each week, creating more work and variability downstream. But in reality, more frequent statistical forecast runs actually cause less change period to period and are easier to react to downstream, making the supply chain more agile and properly balanced overall. By running stat weekly, you can pick up on trends more quickly and give yourself the potential to increase purchasing or manufacturing as needed. In contrast, generating a new stat forecast only once per month reduces the time available for supply planners to react to real sales and market changes.
Data Points
When considering how much historical data to use, it is generally accepted that the more data you have, the better the statistical result will be. Statistical algorithms are able to model patterns and volumes better when they have more data points to build upon. Identifying seasonality and trends is also more effective and accurate when you have more data and multiple seasons to pull from, allowing you to determine if seasonal spikes or valleys occur in the same periods year over year. Sometimes a pattern or volume from several years ago is not reflective of recent market activity—maybe you added or lost a customer or implemented a price change—and short-term demand could look differently. By analyzing the short to long term volume and patterns and using an adequate amount of historical data to properly capture seasonality, you will better predict current demand volume.
Models
Forecasting applications like SAP IBP and APO offer multiple algorithms for statistical forecasting. Some of the algorithms are basic and easier to understand; others can be much more complex and employ artificial intelligence or auto-adjust to consider demand and provide the best output. I typically suggest starting with the more basic algorithms to consider your trends, seasonality and volume. Using a ‘best fit’ function, as mentioned above, is great for determining the most commonly used algorithm for a forecast combination. Be mindful though that a ‘best fit’ may select a different model each time you run stat if irregular patterns or anomalies occur in the demand history. Conducting a proper analysis to find the natural historical pattern is the best way to define which models to apply, avoiding random ‘overfitting’ of incorrect algorithms.
Measure
If you have a baseline statistical forecast, take a snapshot, or lag copy, that you can measure the accuracy. How? Simply set up a periodic copy/snapshot job to copy the statistical forecast, and at month end, calculate the forecast accuracy (FA) of the statistical forecast. You can then compare the stat FA to the final demand or consensus FA. You might even find that your baseline statistical forecast is actually better than your manipulated consensus forecast. If it is not, or is not very close, then you should further dig into the statistical settings/parameters and make adjustments as necessary.
Revisit
Once you enable statistical forecasting, avoid the common ‘set it and forget it’ mistake. If you simply let it run forever and never check the performance result, your outcomes will slowly fall out of line. Your demand will very likely change over time, which means the forecasting patterns and volumes will change. A new product will gain more historical data points, whereas a long lifecycle product may lose customers. Quarterly, semi-annual, or annual analysis is recommended to ensure you are forecasting the right amount of demand, using the right algorithms and proper amount of history. Revisit your statistical forecasting to keep your supply chain accurate.
Final Thoughts
Statistical forecasting is a great way to create an ‘easy to compute’ baseline forecast. The baseline statistical forecast can then be reviewed by the various planners (and even the sales, marketing and finance teams) and adjusted as necessary. If the stat forecast can be captured and stored, you can compare the stat FA against other input forecasts. Performing some analysis to find the proper patterns, time buckets and levels of aggregation will assist in applying the right algorithm.
And finally, don’t set it and forget it! Start strong and start with a solid stat!
For additional resources on forecasting, reference our free archive videos from the webinar series titled Winning Strategies for Statistical Forecasting.
