Improving forecasts through advanced data science approaches

Many businesses use traditional statistical forecasting methods in demand planning processes, and they’re effective for most basic needs. But there are more advanced data science approaches that can improve the accuracy of forecasts and provide additional insight into their drivers.

When is the right time to go beyond traditional univariate time-series statistical forecasting methods?

Data availability and process simplicity are the main reasons to stay with traditional univariate models. In simple terms, you only need time series data from historical orders (or shipments) to generate the forecast. The data is typically easy to source from an enterprise resource planning system (ERP) or similar transaction system. This data is of relatively good quality as it undergoes stringent validation processes as required for use in financial and other legal situations. Most of the following advanced approaches require additional data series to be continuously available, and they need to be robust and validated to avoid skewed or otherwise sub-optimal forecast accuracy results.

If you’re confident that acquiring new data elements won’t become a major obstacle, or that you can invest in developing or purchasing them, then you’re ready to explore outside traditional univariate forecasting methods.

Adjusting forecasts as post-processing

One of the simplest enhancements to a univariate forecast is to adjust it with something you may know (or predict) about the future in a post-processing phase. Typically, this means adjusting the original forecast with one or more factors. If you use multiple factors, they can have different adjustment weights. These factors can include:

Macroeconomic indicators

1. Commodity price indices (copper, steel, coal)
2. Economic activity indices (freight volume, building permits, zoning, gross domestic product (GDP), new mines opening)
3. Economic confidence indices purchase managers index (PMI) and various consumer confidence and sentiment indices

Market intelligence data

1. Actions of competitors (pricing adjustments, footprint changes)
2. Market research (trends, product positioning changes, relative product ratings, demographic shifts, market size estimations)

Internal signals

1. Financial forecasts (What is the confidence within the client to commit?)
2. CRM pipeline value (trending monetary value)
3. CRM pipeline activity (trending opportunities)
4. Front-end sales expectations

It’s often a good idea to do a cross-correlation study before adjusting the forecast with weighted indices or other factors. Usually, you’ll find the greatest benefit from macro-economic indicators like PMI in post-cyclical manufacturing industries, whereas consumer confidence or sentiment metrics can add value for fast-moving consumer goods (FMCG) concerns. Selected factors need to be normalized, weighted and then applied to the forecast using percentage shift per time bucket. You should automate gathering, normalizing, weighting and applying these factors using tools like Alteryx or KNIME.

Multivariate forecasting

In multivariate forecasting, multiple historical data series are used as input to the forecasting algorithm.

Each time series has its own history, which affects its future predictions, but these multiple time series also have a dynamic between them. Although this may sound beneficial, their business forecasting use is rare, as multivariate algorithms can be difficult to understand and explain. They have an unfortunate “black-box” quality in the eyes of typical business users. Also, end results with traditional multivariate methods don’t provide high-enough levels of added value to justify their added complexity. Artificial intelligence and machine learning (AI/ML) approaches have completely altered this equation, and we’ll discuss that later in this post.

Composite forecasting

Composite forecasting is a method of blending various forecasts into one final forecast. In business forecasting, these forecasts are often done at multiple dimensional aggregation levels, AKA “top-down and bottom-up” (separate or multiple dimensions combined into various multi-dimensional segments).

1. Product: SKU – Sub Family – Product Family – Product Line
2. Customer: Customer – Customer Group or Channel
3. Geography: Region – Country – Continent
4. Time: Week – Month – Quarterly – Yearly

Lower, more detailed forecasts bring individualized pattern signals into the forecast, and higher levels burnish exceptions or outliers. You can also identify seasonal patterns at higher levels for items having abbreviated sales histories where such a pattern is not yet clear. You can do a composite forecast by grouping lowest-level historical time series data to various “dimensional segments,” generating forecasts for all of them, and then blending them together with a chosen weighting. This extra preparation must happen outside the core forecasting solution using a data science tool. However, there are automation methods like thief-library in R for automatic temporal aggregation/disaggregation. Also, current data science platforms such as Dataiku, Alteryx and KNIME make this much easier to implement and automate than old-school, hand-crafted SQL or Python solutions.

AI/ML time-series forecasting

The new Shangri-La for forecasting? Not quite, but there are many interesting possibilities. In recent years, there’s been rapid development in AI/ML for time series forecasting. Many business AI/ML use cases focus on classification challenges, which often have striking results. Major investments in AI/ML platforms, driven by venture capital money, have made these solutions more useful in everyday business settings. Best-in-class AI/ML platforms like DataRobot and H20.ai automate most steps of ML model training, validation and explanation (although access to every single detailed parameter still exists for the data science-oriented employee) and getting concrete results has become as easy as when using any other normal business software. AI/ML has “escaped” from the domain of academics or the rare data science Ph.D. to becoming a daily tool for demand planners and business analysts.

Although the modern ML algorithms often underperform traditional univariate methods (although we are seeing continuing improvement in real-world cases), ML really shines when you have additional data to increase the predictive signal. One of the significant benefits of automated ML solutions is that you don’t have to know any correlations, weightings or relations about the additional data beforehand, but the machine will find them and utilize them accordingly.

With recent real-life use case piloting, we have seen amazing results with AI/ML classification cases like “will this CRM opportunity convert into an order?” For many divisions of a large global company with several distinct sales patterns, we can reach 80-90% accuracy scores for the predictions. These are major drivers to the future demand plan along with the statistical forecast for this client.

In other words, during prototyping you can incorporate various time-dimension information into ML learning data and see what increased forecast accuracy and what didn’t. ML also automatically “weights” the importance of various data elements (or “features” in ML parlance).

AI/ML-based time series forecasting is often also called “ML driver-based forecasting” where the drivers are separate time series (like the ones mentioned above in the “Adjusting the Forecast” paragraph) with varying weights towards the final forecast. Modern advanced planning tools have these capabilities inbuilt with rich visualizations to avoid any ‘black box’ obstacles during implementation.

NTT DATA offers AI/ML introduction and use case piloting services (a low cost, no-investment approach to determine if AI/ML brings true value), Machine Learning-as-a-Service (MLaaS) and general project-based consulting around Advanced Analytics and AI/ML domains.

— By Sami Salminen