FORECASTING IN FLUX: DECODING THE DYNAMIC DANCE OF PRICE CHANGES AND DEMAND PREDICTIONS
In the highly evolving and dynamic world of today’s market, where disruptions have become usual rather than the exception, understanding and predicting customer behaviors in response to external factors are crucial. The COVID-19 pandemic has disrupted supply chains since the lockdowns have led to decreases in production and the logistics problems have caused longer disruptions in many sectors. Eventually these disruptions caused persistent inflationary pressures. Due to these external factors, traditional forecasting models with limited features fall short, prompting further investigations regarding understanding customer behaviors of today’s unsteady markets.
These investigations are implemented with the Continuous AI approach Koçdigital embraces. For a deeper exploration of our Continuous AI approach, you can visit: A Manifest for Continuous AI. Digging deeper into why traditional demand forecasting models fail, three factors emerge: data drift, concept drift and algorithm drift. For detailed explanations of these factors, you can check our blog post: Drifting Effects on AI Models: Why Continuous AI is a Must?
Among the numerous variables contributing data drift, price volatility plays a significant role. It is observed that sales of regular customers fluctuate especially during price list change periods, since the companies unveil the alterations in price in advance and these announcements trigger shifts of usual orders. In a highly inflated environment, the most common price list change is price increases. Some of the regular customers may stock up on their purchase orders depending on their stocking capabilities and cash flows which eventually affects the demands for products. These cases are often seen in fuel-oil, fast fashion products, cosmetics and other sectors that announce their price list changes officially or word-of-mouth. Good news is that these behaviors are interpretable, thus they can be foreseen. In this blog post, I will explain how to enhance forecasting models with comprehensive studies that account for price volatility.
Who disrupts the predictions?
First, we need to make an analysis to determine the scope of the study. To foresee the customers who can buy excessive number of products and stock up should be detected as not all customers are able to do so, due to two main reasons: They may not have enough cash to buy more than their regular orders before the prices increase or the limited stocking (and protecting) capacity of customers do not allow such strategies. This kind of stockpiling requires idle storage and thus only a limited number of customers may have enough space and equipment to store and protect the products properly. Once these specific customers are detected, their orders during these price list change periods should be analyzed.
When does the accuracy of predictions decrease?
In highly inflated countries, pricelist changes occur quite often, though not all price changes create the same fluctuations at customer side. The case can be limited to only a subset of price changes that are relatively higher than the rest. Customers do not bother for minor price changes as those will be seen not worth tying the cash flow to the product purchases of future periods and/or occupying the limited warehouse space.
What happens as prices increase and what can be done to predict more precisely?
Since the “who” and “when” are determined, now “what” should be found. The sales to “the customers” during the specific pricelist change periods should be analyzed. When the price increases are announced in advance, the regular customers which are specified before, advance their order. This behavior leads to sales increase in the short term just before the price increase. Since these customers meet their demands in advance, they postpone their next orders. That’s why sales decrease after the price increase. Eventually when the balance is achieved, sales are back to their previous amount.
To elaborate on this case, let’s illustrate it with an example. Let’s say that a distributor sells coffee beans to coffee shops weekly. Due to high maintaining costs, the distributor has decided to increase the price. Previously, he used to sell coffee beans for 5.00 $ per kg. The new listing price is 7.50 $ per kg. The raise ratio is %50, which can be considered high. This price increase may lead to customer loss but let’s assume that this market is not highly competitive, and the customers are loyal to the distributor. This distributor announced the price list change beforehand to provide transparency. Some of the customers decided to rearrange their orders and increased them, at the same time as they hear about the coming change. Hence, sales right before the new price arrangement were increased. The next week, since the customers have already enough coffee to cover the demand for a while, they would not order coffee as usual. In summary, coffee demand fluctuated during these 2 weeks. The graph of these sales would be similar to the graph shown below. Week 5 is the week in which the new price is valid. Week 4 is the week in which the new price listing is announced. The model accuracy of both 2 weeks decrease.
As can be seen in graph 1, the sales fluctuate between week 4 and week 5. The total sales of Week 4 and Week 5 are 170. Total prediction for these 2 weeks is 152. Total accuracy seems fine with %11 mean absolute percentage error. However, weekly accuracy of both 2 weeks is low. As we analyze the sales further, we can see that some of the sales of week 5 shifted to week 4. Hence, we can say that a %50 price change has led to a %15 of total sales shift to former week, creating imbalance. If we could split our total prediction for these weeks with the ratio 65:35, the new calculated accuracy would be higher than the previous accuracy. The new graph of adjusted predictions and sales can be seen below in Graph 2. Of course, to come up with an accurate ratio, many historic price adjustments should be analyzed. If the total accuracy of these two weeks seems successful, shifting some of the orders from the second week to the first week can improve the accuracy of the model.
Accurate price change data for future periods are critical to apply these kinds of adjustments to the base predictions. As the behaviors of customers are linked to the price change expectations, this data should be updated regularly. The more accurate the price change expectation data is, the more precise the predictions will be.
What if the opposite happens?
The exact opposite of the mentioned customer behaviors can be observed in price discounts. When the customers hear that a special discount would happen or a new promotion /campaign will be run in advance, they may tend to postpone their non-urgent demands. Again, not all customers can postpone their orders, since they may not have enough stock to handle until the start of promotion / discount date and/or not all the discounts are considered worth postponing the orders and facing stock out risks. The first step is to analyze the customers who would be able to postpone, and discount ranges they would be tempted to do so. Next, the data should be analyzed using previous discounts and/or promotions, to understand which and how discounts /promotions would affect demand. Only after a clear pattern is seen the predictions can be improved with some adjustments.
Let’s assume that a fashion brand announced that it would apply a %50 discount at its products 1 week in advance. Customers postpone their orders to the week in which the discount would be implemented so that they can make their orders at discounted prices. Graph 3 represents an example of discount announcement effects on demand. Week 4 is in which the announcement is made and week 5 is in which the discount is implemented. The ratio of sales of week 4 to the total sales of week 4 and week 5 is %40. Since the accuracy of the total prediction seems reasonable, we can use this ratio to find out the adjusted prediction of week 4 and week 5. The adjusted prediction with this ratio is represented by the green line.
Conclusion
In unsteady and dynamic market settings, it is crucial for companies to predict uncertainties to align strategically. Customer behaviors in such settings might set off a chain-reacting effect in supply chains. It is critical to understand what are the triggers for distinct customer preferences. In environments marked by frequent inflation, price increase announcements can cause stockpiling before increase takes place, eventually postponing some of the orders to later periods. Price decreases and promotion/campaign implementations may also lead to fluctuations by decreases in sales prior to these implementations and increase in sales after. Similar patterns are generally repeated in similar price list changes. Hence, if the pattern is understood and interpreted clearly, predictions can be adjusted more accurately.
