Demand Forecasting is one of the major aspects of any product based organisations. Accurate forecasts lead to production efficiencies and better inventory management and prevent cost overruns. AI based models have been really instrumental in improving accuracy and transparency of this function. There are two types of models.

Black Box Models where Forecasting and the explanation of the model are decoupled. Explanation of the model increases transparency and hence extremely critical for understanding. The researchers here develop separate algorithms for to understand the model.

Glass Box Models where the AI based algorithm can explain its prediction.

Forecasting planners require explanation about the factors and data that went into each forecast to better understand the soundness of the forecast and thereby has more control. An explanation of a model should have a meaningful information and have a logical explanation (Pedreschi et al. (2018). The model should also have actionability information and may also provide counterfactuals. The technology and the algorithm used may have the following features.

Provide knowledge about the features of the dataset.

The characteristics of the dataset and the ranges.

Provide explanations in the correct language and context for the user.

Let us now understand why AI and Machine Learning based models are preferred.

Traditional versus Machine Learning based Forecasting Techniques.

Traditional statistical forecasting models are based on historical time series data and their accuracy is dependent on stable market conditions. However traditional methods fail to take into account changes in consumer behaviours and disruptions in the marketplace due to innovation and technological breakthroughs.

Machine learning algorithms can use internal and external real time data to create much more relevant and timely forecasts. Institute of Business Forecasting and Planning outlines certain datasets that modern machine learning algorithms can use. They include sales data, website statistics, clickstream data, geolocation, macroeconomic indicators, social media, POS data, third party syndicated data etc.

Machine learning algorithms uses such large datasets and complex interrelationships among various datasets to understand patterns and capture demand. The smart models continue to retrain themselves over time as fresh data comes in and thus addresses volatility of the market. This gives much more accurate and reliable forecasts.

However, as we discussed earlier in the chapter, it is imperative for a human understanding of the features of the model, the impact of each dataset on the outcome and also the rationale for the forecast.

Predictive Analytics

Predictive sales analytics is the most common application of machine learning algorithms along with statistical techniques. It helps the company to understand demand and also consumer buying behaviour in certain scenarios.

The predictive models combine internal company data with external macroeconomic data and many other external variables to predict demand. It helps an enterprise to take much more informed decision about consumer buying, product launches and planning for an event.

One of the major drawbacks is it can at best look at medium term forecasting because of the complexities of the datasets involved. For more near understanding of demand, techniques like Demand Sensing is used.

Demand Sensing

Demand Sensing algorithms incorporates real time sales data to throw up short time forecasts as variabilities in buying appears. It extracts daily data from sales, warehouses, and other sources to indicate fluctuations in the medium-term forecasts. The models also explain the causes of each variability, significance of each factors and offers very short time forecasts for day-to-day functioning.

Popular Deployment of Machine Learning in Demand Forecasting

Since machine learning algorithms are costly to deploy and requires huge amount of data and computing power it is essential to understand how an enterprise can extract the best possible value from this practice. Some best use cases are.

New Product Introduction: New product demand forecasting is difficult due to lack of past sales data. Popular methods of such forecasts have been the traditional expert opinion based on experience. However, machine learning algorithms use past sales data of similar products and product life cycle curves to do a more accurate and data based forecasting. This impacts decision making for production, marketing and supply chain functions of an enterprise.

Product with Short Life Cycles: Fashion products are typically seasonal and the life cycle is about a few months. In such a scenario, Demand Sensing tools are of great value since it gives a sense of predictability and control in the entire process.

Seasonal Products: Products like winter garments, umbrellas and summer products like cold beverages are dependant on the weather conditions and require extreme short term forecasts. Machine learning helps enterprises to build various scenarios and the effect of each on the demand.

Practical Illustrations

Global brand Luxottica, world’s largest eyewear company introduces around 2000 new styles every year. The use of smart forecasting techniques have helped the company to improve its forecasts by 10% resulting in significant savings and control.

UK National health system uses ML based algorithms to understand the blood requirement in hospitals to prevent overstocking and stockouts. Since blood is a critical live saving item, no stockouts can have very disastrous consequence and overstocking leads to wastage it is very important to have accurate and reliable forecasts. This method has reduced overstocking in hospitals by 30% resulting in less wastage at some hospitals and stockout in others.

Demand Forecasting Methods: Using Machine Learning and Predictive Analytics to See the Future of Sales.