Using Machine Learning Forecasting for Better Demand Planning
Leverage more data and achieve dramatic accuracy improvements at enterprise scale with HaloBoost©
Forecasting is becoming more complex, with many firms striving to incorporate product, pricing, discounts, channel, and other available data to improve accuracy. This increase in forecasting complexity and the associated massive increase in data volume requires a Machine Learning (ML) forecasting solution. Simply stated, traditional forecasting methods cannot scale to the massive data and SKU level forecasting that clients demand. That is why Halo has released HaloBoost©, the first of its kind Machine Learning system for demand forecasting. Tested and proven on dozens of large databases, HaloBoost© is simple to implement and
is a powerful new tool for your planners. These ML forecasting solutions are aligned with the basic Halo architecture, allowing Halo customers to test into these ML solutions, see the proven accuracy gains, and then adopt ML forecasting on an evidence-and-value-add basis. In addition, ML forecasting is very fast, allowing a company to generate hundreds of thousands of SKU-level forecasts in minutes. And with Halo’s dashboard and report management services you can get your ML forecasting results into action quickly because the Halo system has been designed for this type of enterprise-scale, massive forecasting business case.
Ready to see Halo in action with your data?
Advantages of Applying Modern Machine Learning to Demand Planning
More Data - More Accuracy
Typical forecasting methods project future sales from past sales levels; seasonality and cyclical trends are incorporated, but price, discounts, product features, and sales channel information are often ignored during forecasting, and accounted for later in adjustments. Machine Learning forecasting allows for more information to be incorporated into the forecast.
The forecast is optimized at the level of the individual SKU, incorporating what is known about pricing history, discounts, and other factors that may be under management control. Product ingredients, packaging, raw material pricing, third-party economic data, and virtually anything that can be measured can be incorporated into the forecast.
Build Forecasts Faster
Building forecasts at the SKU level may appear to be CPU intensive, and there is some truth to that. But ML computer algorithms have been developed capitalizing on the parallel processing capabilities of modern computers, resulting in lightning-fast forecast generation.
Benchmarking has demonstrated the capacity to build more than 1 million forecasts in an hour, with no sacrifice of accuracy, using commodity-priced hardware that is easy to procure and set up.
Deep Insight Into Your Forecasts
ML forecasting methods can appear to be a “black box” where even a more accurate forecast is viewed with skepticism when the complexity of the forecasting model defies simple explanation. Halo has focused on ML methods that lend themselves to interpretation and the generation of value-added insights.
The output of ML forecasting documents the relative importance of various data sources; data importance insights improve interpretation and provide feedback on what data can add value and should be curated for future use, versus data that does not improve prediction and thus can be archived for lower storage costs.
HaloBoost© Scores Top 95 at the Kaggle.com Predictive Modeling Competition
In a recent competition for Predictive Modeling hosted by Kaggle.com, HaloBoost© scored in the top 95th percentile in the world with our fully automated and production ready modeling approach to retail sales forecasting.
How HaloBoost© Works
Learn How HaloBoost© Transforms the Demand Planning Process
Many use cases can be characterized by a product mix where 90% of sales volume is accounted for by 20% of the products. By segmenting on volume, price, and frequency of sale, a massive forecasting space across hundreds of thousands of SKUs can be broken down into immediate high value opportunity, marginal opportunity worth pursuing, and space where SKU level forecasting is not practical due to sparse sales volume and limited sales history.
Halo’s forecasting solution builds this segmentation step into the workflow at an early stage, so that forecasting can progress most rapidly on the opportunity that is most valuable; once initial forecasts are proven accurate and valuable, marginal segments can be included until the diminishing returns are reached. All remaining SKUs can still be forecasted, either individually or in aggregated segments, depending on business needs.
Two-Stage ML Forecasting
HaloBoost© is Halo’s proprietary method of “stacking” machine learning algorithms to yield results fast, and then fine tune for accuracy and bias reduction. Using two dominant ML algorithms rather than one approach allows Halo to ensure the forecasting results are robust and not “method-specific”, reducing the risk of future forecast bias.
Stage 1 of HaloBoost© improves accuracy by 25% in most use cases, with Stage 2 adding incremental time (4 hours for 1MM SKU-level forecasts rather than 1 hour) but often with an additional 67% improvement in accuracy. Review of results across Stages 1 and 2 then allows us to tune HaloBoost© for production use.
We use industry standard accuracy metric and can code custom accuracy metrics on client requests.
The Halo dashboards then facilitate drill-down into the validation to identify any segments where accuracy is sub-optimal and where further data exploration and ML tuning may be beneficial.
Monthly forecast accuracy monitoring ensures that the massive data used in ML forecasting is stable and not introducing bias. Halo has developed month-over-month forecast monitoring dashboards that document and trend accuracy metrics in great detail.
Variances from forecast that approach tolerance limits are spotted earlier so that corrective action can be taken before a forecasting variance impacts business results.