Meet Julie! She sells ice cream on the Moon
This is Julie. Since the age of six, all she wanted to do is to sell yummy ice cream and bring joy to people every summer. She believed she could, so she did.
Julie is now the boss behind Salty & Icy, one of the best ice cream brands on the Moon. What started as a small space shuttle is now distributing millions of ice cream tubs to space travellers who were eager to explore the craters, fell in love with the beautiful lunar landscape and eventually settled down on the Moon like Julie.
Sounds like a dream, but there is more to running a business than knowing how to make delicious ice cream!
Making ice cream, then delivering it to stores all over the Moon takes a lot of time. Yet when customers want ice cream, they won’t settle for more than a 10-minute wait or any other flavours except for the one they are craving.
Simply put, ice cream has to be made well in advance before customers even know they wanna buy ice cream. But how much should Julie produce? 1000 or 2000 tubs per day? Chocolate or strawberry or 20+ exotic flavours that Salty & Icy is famous for? Single-serving cup or 16-ounce container?
Producing too much of what customers don’t want to buy will see Salty & Icy losing money because products that aren’t sold within two weeks would have to be thrown away. Producing too little of what’s in high demand will force her precious customers to opt for other competitors and lose them forever.
Believe it or not, the mission all boils down to one question: How much of each product that customers are likely to buy in the future? Finding the answer to this question is called demand forecasting.
Demand forecast empowers Julie with better business planning and decision-making
Generally speaking, here is why demand forecasting matters so much for Salty & Icy (as well as other businesses).
Being able to predict demand within the next 2 weeks could help the Salty & Icy team mitigate stockouts while timely execute promotions and markdowns to drive sales.
A slightly longer forecast horizon of up to 2 months could enable better production planning to keep the ice cream churning to meet customers’ demand.
Beyond 2 months, a reasonable view of forecasted demand could guide Julie and her team towards better product development and business strategies for future growth.
Looking at a broader picture, the longer the forecast horizon, the further into the future Julie can plan, schedule and provide for the raw materials, labour and production capacity to fulfil customers’ demand and execute growth strategies.
If only Julie has a crystal ball to tell her what the future business holds!
Patterns & Unexplained Variance: 2 Components of Demand Forecasting
Without all of the glorious statistical methods or AI hype, the only goal of demand forecasting is to predict the quantity of products or services to be consumed in the future based on existing data. That’s it!
The more data about what has been influencing demand for a product from the past till now, the easier it is to spot the underlying patterns, relationships and create a decent forecast about what the future business might hold.
But what forms a demand forecast? Here is a short and sweet summary on what Charles W. Chase Jr. wrote in his book.
Patterns often come in either of the following 3 forms: trend, seasonality and cyclical factors.
- Trend: Since historical demand for vegan ice cream has been increasing by x% annually for the past five years, Julie can expect to sell x% more vegan ice cream in the future.
- Seasonality: When the sun hits the moon’s surface, the temperature rises. That’s when space travellers go crazy for ice cream to relieve themselves from the scorching heat. Vice versa, when the sun goes down, temperature can dip way below zero. Very few people fancy a brain freeze from an ice cream tub.
- Cyclical: Daytime and extreme heat on one side of the moon lasts about 13.5 days, followed by 13.5 nights of darkness and freezing temperatures. Demand for ice cream for each side of the moon also rises and falls in tandem with the lunar rotation.
The underlying assumption here is what customers want to purchase in the future will mimic what they used to buy in the past. So if Julie can identify the trend, seasonality or cyclinal factors that has been influencing demand for her products in the past, she can predict the future demand.
Sounds too good to be true?
I agree. The future won’t likely behave the same as the past. And that’s where the second component comes into the picture.
Unexplained variance represents the discrepancy between historical demand and future demand that is not explained by either trend, seasonality or cyclical factors.
But what might this be? Hmm, many many different things.
This could be a buy-1-get-1-free promotion or a 5% price reduction that other ice cream brands are doing to compete with Salty & Icy. The Interstellar Cup hosted on the Moon could see a rise in demand for large-size multi-flavoured ice cream tubs since friends and families tend to watch soccer matches together. An unfortunate accident related to the space shuttle could make people become wary of space travelling, thus affecting the overall trade and economy on the Moon.
Each of these phenomena affects demand for Julie’s ice cream to varying extents. And only God knows how many more factors that could drive sales up or down.
But because it’s tough to predict the future with 100% accuracy doesn’t mean Julie should abandon it altogether. In fact, the aim and the expectation should always be quantifying unexplained variance based on data as much as possible, within reasonable time and cost contrainsts.
But how? Well, let’s explore what people have been doing all the time to make a reasonably accurate demand forecast a reality!
There are more than one approach to predict future demand
I have no intention to write a laundry list of all forecasting methods under the sun. This would put you straight into a deep sleep.
But let me briefly introduce 3 broad categories of demand forecasting methods. Each category represents an underlying belief or behaviour from people who are running the business in real life that is worth knowing. So hear me out!
Time Series: History will repeat itself
Have you ever heard of moving average? How about exponential smoothing? Oh, how can I omit the almighty ARIMA model for time series forecasting?
If you need a quick recap on what those are, check out this article. But below is the most important point you need to know.
Using these statistical methods, people intentionally or unintentionally assume history will repeat itself. Hence extrapolating the historical data with time-series modelling seems to be sufficient to forecast demand.
For instance, the quantity of ice cream Julie sold this month would indicate the demand for Salty & Icy products next month. The cyclical fluctuations in demand for ice cream based on lunar rotation or the rising popularity of vegan ice cream that Julie observed since launching her business would tell her how much ice cream she should produce next quarter, next year and many years to come.
But is it true that the past history would remain constant over months or even weeks? Let’s see.
Personal Judgment: Trust your gut
Have you ever seen how a baseline forecast is circulated around a committee to gather feedback and approval?
Every quarter, Julie often prepares a baseline forecast based on time series modelling. She then sends the baseline forecast to the entire management team.
Each person adds his or her own twists to the forecast based on judgments, experience or gut feelings because everyone knows history will not repeat itself in its entirety. Some would happily adjust the forecasted demand upwards while others seem to be too pessimistic about the demand in upcoming months.
So who is right? Who is wrong? And more importantly, why do people choose to rely on gut feelings anyway?
Many researchers have investigated this interesting behaviour. To keep it simple, here are 3 main beliefs that might explain what’s happening on the ground.
When the time to make decisions is pressing and the time series forecast is either not available or not robust enough to capture the sophisticated market behaviour, people would rely on domain knowledge, personal judgment or a little bit of both.
Unfortunately, regardless of whether the judgment is done by one executive or an entire committee, there is a very fine line between domain knowledge and personal bias. Research has shown that most of the time, adjustments based on gut feelings can produce highly inaccurate forecasts when the figures reflect someone’s wishes or flawed reasoning instead of a genuine understanding of market and customer behaviour.
Hmm, there must be another way, right?
Regression: Above all else, show me the data
Beyond trend, seasonality and cyclical factors, what else could help to explain away any unexplained variance related to future demand? As Julie dug deeper into the past records, she observed several relationships.
- Spending more money to display her products at eye level in the centre of the frozen aisle at supermarkets, sales volume would be twice as high as other areas.
- During the “frozen price war” last May, most ice cream brands on the Moon were on sale for half of the price of Salty & Icy. Sales volume dipped by 25%.
Although correlation does not imply causation, the interrelationship between external and internal factors could be used to forecast future demand. For example, hot weather correlates with 5% higher than average sales volume while a price 5% higher than other brands might cost the business 20% of weekly demand.
Quantifying such relationships and defining a mathematical function to predict future demand based on what is known about those factors at the moment is known as regression.
Remember we spoke about 2 components of demand forecasting? While time series methods help to uncover patterns related to time, regression is extremely valuable in quantifying unexplained variance.
Bringing our discussion full circle to Julie and her needs to predict future demand for Salty & Icy products, here are what she needs to bear in mind.
Firstly, making those judgmental adjustments to improve the accuracy of a quantitative forecasting model is a no-no. A better approach is to validate the domain knowledge with existing data via hypothesis testing.
Secondly, time series methods where history will repeat itself will only work if the products have a fairly stable sales volume over the years, with very small random fluctuations. Nevertheless, since the marketplace is moving so fast and new entrants crop up everywhere, very few firms can continue to enjoy stable sales volume.
Thirdly, the more data representing various influencing factors to product demand from different sources, the greater the possibility to improve business insights and forecast accuracy. But doing regression analysis with many potential factors could quickly become a nightmare with Excel.
Finally, beyond the AI hype, it takes more than Machine Learning to generate an accurate demand forecast that can be used for decision-making because:
- Machine Learning doesn’t solve data problems. If historical data is missing or has poor quality, Machine Learning alone can’t generate a demand forecast model with precision and accuracy.
- When people don’t trust the accuracy of the forecasted figures or don’t understand how the numbers come about, it’s difficult to convince them to rely a little less on gut feelings and a little more on data. Therefore, a Machine Learning journey would probably take time, lots of pilot projects and cultural shifts to arrive that the sweet spot between an “accurate enough” model and an “explainable” model for understanding and adoption.
That’s all I have for this blog post. Thank you for reading. Have feedback on how I can do better or just wanna chat? Let me know in the comments or find me on LinkedIn. Have a fabulous week ahead, everybody!
- Do ‘big losses’ in judgmental adjustments to statistical forecasts affect experts’ behaviour? by Petropoulos et al.
- Demand-Driven Forecasting: A Structured Approach to Forecasting by Charles W. Chase Jr.
- Demand Forecasting for Inventory Control by Nick T. Thomopoulos