This is a high level introduction to demand forecasting, for those who have only recently heard about this powerful offering, and wish to learn more.
Demand Forecasting refers to suite of analytical tools and methods that estimate the demand for both the inputs you require to do business and the outputs that make up your business — be they services or products. Traditionally, these estimates are based on historical data — seasonality and trends. As technology and the statistical methodologies become more advanced, so does the accuracy.
The new methods take into account variables such as public holidays, competition, market forces, marketing efforts, the weather, events within a region, and are also highly varied to account for the different demand patterns for different goods and services.
Algorithms process data and produce an output, plus a degree of confidence. These algorithms vary in complexity, applicability and use. As such, it can be advisable to apply different models to your data to determine which model is best, but be aware that different products and services will have different demand patterns.
Finding Patterns
When it comes to predicting future events, the ability to unpack the data that you already have and look for patterns and changes in behaviour is crucial. To illustrate simply, we’ll take a look at seasonality and trend.
Seasonality follows a pattern that repeats at regular intervals (such as sales in retail increasing around Christmas) Trends, on the other hand, describes the direction that data moves in over a period of time. In retail, you might see a trend of increasing sales followed by a decline or plateau once a good is no longer fashionable. where the different directions (and change in direction) can also be described as trends. Seasonality and trend often appear together, and you’ll need historical data in order to identify them.
In order to create forecasts that capture these patterns — thus enabling you to capitalise on them — there are several tools and methods to choose from, depending on the data at your disposal and the purpose of your forecast.
Types
Whether your data has seasonality, a trend or two, or a variety of variables, there are two tools that can take these characteristics into account when creating forecasts: traditional and machine learning methods.
Traditional methods, such as ARIMA and Exponential Smoothing are well-suited to
Short-term forecasting
Data that has stationary properties (such as seasonality and trends)
Limited variables that affect demand
Machine learning algorithms, on the other hand, are suited for creating forecasts when there are lots of variables that influence demand. These algorithms can incorporate the effect of these variables — think everything from marketing efforts, weather, exchange rates, web traffic and even large scale events such as a football final — into the forecasting model. These characteristics are part of the reason why Machine Learning methods create forecasts with high accuracy — up to 50% more accurate than traditional methods. But, the main drawback is that they can need large amounts of data in order to work.
If your data doesn’t quite fit any particular method — or you’re unsure which one is right for you — you can build an ensemble model. This means that you combine different algorithms which all run predictive models in simulation and agree on which result is best suited at an SKU (stock keeping unit) level. This is the best approach to generating the greatest accuracy at a microscopic level, to ensure maximum efficiency across all stock.
What’s It Good For?
Demand Forecasting can have different uses that depend on the objectives you have, the time period that you want to forecast for (the further out, the harder to attain accuracy), and the complexity of the data. In terms of objectives, Demand Forecasting can be used in the short-term to predict sales, better prepare you for seasonal changes, and formulate production and price policies. In the mid- to -long term — Demand Forecasting can be used in planning for sales and marketing, finance, and capacity, and it can be the driver behind business strategies.
For example, retail, tourism, and other industries use Demand Forecasting to optimise pricing of products, manage stock, and plan marketing campaigns based on which products are predicted to sell well. In comparison, industrial manufacturing has also seen the benefits of Demand Forecasting, whether it’s predicting how much of a unit to produce and be made available or the best times to perform maintenance.
Challenges
There are some challenges that accompany Demand Forecasting. Difficulties arise when forecasts are frequently created for lots of products. This results in an overload of information that needs to be navigated quickly in order for it to be meaningful and useful, which is sadly not the case and thus becomes a waste of time and money for both the people generating the forecasts, and those who go through the results but don’t act on them in a meaningful way. This can be mitigated by integrating AI strategies (such as AI auto replenishment) to increase automation. Additionally, some methods can also run into issues when dealing with expensive, low-velocity products such as white goods and cars — meaning that less than one item is sold per week. This means that, when you create a weekly or monthly forecast, these products can be predicted as zero and you may need specific algorithms in order to get a more accurate result. There are creative ways to get around that, but we won’t reveal that just yet.
Final Thoughts
Keen to learn more about Demand Forecasting? Check out some of our case studies (there’s plenty to choose from), where you’ll find out how we’ve used Demand Forecasting to improve the forecasting approaches used by a Fortune 500 company, an Australian consumer wholesaler, a 100-year-old manufacturer, and the retail arm of a Fortune 500 manufacturing firm.
And, if you like what you see, why not check out our blog or sign up to our Recommended Reading newsletter here for the latest AI reads.