Accurate demand forecasting is essential for any inventory led business. Improving demand forecasting accuracy helps businesses achieve operational efficiency by optimizing inventory in real time, reducing wastages and increasing margins. 

However, we need to understand that demand forecasting is an ongoing learning process that needs to keep improving over time. In this blog, we discuss the top 5 lessons we learnt from working with a wide spectrum of clients from various industries.

‍

Top 5 demand forecasting lessons

‍‍

Selling at discounted prices doesn’t always yield higher sales

The common notion with businesses is that a discounted price on a SKU is going to yield more demand for that SKU. While this can be true in some cases, demand is not purely tied to price often. There are several other external factors that influence demand in real time. 

One of our clients who is an E-commerce Pet Food company ran a price discount of 5% on their  “Dog Treats” and they saw a whooping increase in demand for that category. This encouraged them to run an increased price discount of 10% on the same category a few months down the line. However, we noticed that the demand did not go up as expected this time despite offering a higher price discount. 

After analyzing this situation, we identified that the website and social media promotions for the 5% discount promotion had better visibility than the 10% discount promotion. This explains the dip in demand the second time in spite of offering a more enticing discount. This shows that even the price elastic SKUs may not always offer a boost in sales given a reduction in prices. It is also important to factor in and quantify other demand drivers including marketing activities, weather, holidays and market trends.

‍‍

Predicting demand at the most granular level may not always give you the best accuracy

In most cases, it is true that predicting demand at the most granular level which is the closest to the customer is the way to achieve high forecasting accuracy. However, when it comes to demand forecasting, there is no one approach that fits all. 

From our experience with a Fashion and Apparel brand, we learnt that predicting at the most granular level may not always fetch us the best results. The client had a wide portfolio of products. Each SKU had several granular combinations across different styles, colors and sizes.  In this case, when we drill down to the most granular level (i.e) SKU x Colour x Size level, the data points get very intermittent or sparse. There is no rich data available for the model to produce highly accurate predictions. 

Here, we set the granularity at one level higher (i.e) at the SKU x Style/Model level which had relatively more data points for predicting demand accurately. In this case, the volume of sales at the SKU x Style/Model level is first stabilized and then the sales at the more granular levels like the color and sizes in each model is predicted. Here, the predictions made at the higher level are distributed proportionately across the granular levels based on historical data. It is also essential to choose the higher granular levels appropriately as this might be different for different business models. 

‍

Richer the external data, higher the prediction accuracy

 Demand forecasting relies heavily on data. Rich data provides detailed information about various factors influencing demand, such as customer demographics, past purchasing behavior, seasonality, economic trends, and market dynamics. This granularity allows for a more nuanced and precise analysis of demand patterns. 

With one of our retail clients, we have been capturing promotional data to train the system to predict promotional uplift in demand accurately. The high level data on just the occasion and duration helped us predict promotional demand accurately to some extent.

However, we saw that feeding more granular level data about a particular promotion enabled us to identify patterns and trends that may not be apparent with limited or basic data sets. For example, a 10% price discount may impact demand differently when offered as an ‘In Cart Promotion’ than as a ‘Holiday Promotion’. This helped us predict promotional demand with unmatched accuracy saving the client from instances of out of stocks or excess inventory even during marketing events or promotions.

‍

When it comes to forecasting, sometimes, a snug fit falls short of the mark

‍Overfitting is when a model fits too closely with its training data. This is a common problem that arises during the data modeling process. In this case, the model fails to perform efficiently with any new data and loses its ability to predict demand accurately. 

For some of our clients, we have seen that the model we have created starts to mimic the training data too closely and is unable to adapt and account for any real word fluctuations. This happens when there isn’t sufficient data for model training. When limited data sets are used, the model fails to identify any patterns or outliers. Another reason for overfitting is due to too much noise in the data. For instance, when we are trying to forecast cyclicity, if there is too much irrelevant data that does not have cyclical trends, then that leads to overfitting.

We have seen that using rich, high quality data without any inconsistencies or anomalies helps avoid overfitting. It is also important to use data which represents the business in the best way. This will help the model to account for market fluctuations.

‍

Historical Data reveals more than what meets the eye

‍While historical data is an absolute necessity for forecasting demand, we often forget that historical data is also a gold mine of valuable insights. Analyzing historical data helps identify customer behavior, demand trends, seasonal variations and potential anomalies. Historical data lays the foundation for demand forecasting by offering visibility into the dynamics of demand.

In our experience with clients across different verticals we have often turned to historical data to make sense of the demand behavior. The assumptions made by the client on the market fluctuations and demand trends can be validated using historical data. This helps in making informed decisions on production capacity, resource allocation and new product introductions. 

It is critical to understand that historical data is more than just another factor impacting demand; it is a treasure trove of useful information. 

In essence, mastering demand forecasting is a journey marked by continuous learning, adaptability, and a keen eye for data-driven insights. By internalizing these lessons, businesses can forge ahead with confidence, optimizing inventory management and driving sustainable growth in dynamic market landscapes.

‍