How can one plan demand in such volatile times?
According to Gartner, a top challenge that businesses face today – irrespective of size, scale, or industry – is demand volatility. Too many factors, including everything from political factors to weather fluctuations and further on to influencers’ posts on social media, impact buyers. As a result, buyers are prone to changing their minds more frequently than ever before.
But what if you could predict demand early enough to plan for, manage, and make the most of it, always? Because the truth is that you can. Not just that, there are many ways to do it. In this article, we look at the capabilities of the most common forecasting methods, and which situations they are best for given their current limitations.
Breaking down demand forecasting
Before getting into the forecasting models, it would be useful to define the three main components of demand forecasting.
Demand forecasting by itself refers to the process of estimating the future demand for a product or service. Though often used interchangeably with the terms demand planning and demand sensing, the three in reality are different processes.
Therefore, demand forecasting is the process by which demand planners predict the demand for a given product in a week’s, month’s, or even a year’s time. Demand planners work with sales, marketing, and other key stakeholder teams to extract historical data such as sales and company growth statistics – as well as real-time data on market trends, consumer behavior, etc. – to run forecasting models and arrive at a solid consensus forecast. This demand forecast in turn forms the basis for the overall demand plan.
The goal of demand forecasting is to accurately predict future demand and is, hence, very data-focused.
Demand sensing is the process of leveraging advanced analytical techniques to create short-term predictions of real-time fluctuations in purchase behavior. This is particularly helpful for modern-day businesses, which function in fast-changing markets.
Demand sensing solutions extract everyday data from POS systems, warehouses, as well as external sources to detect variances in sales by comparing with historical trends. The system automatically analyzes the significance of each divergence, evaluates influencing factors, and recommends adjustments to short-term plans.
The goal of demand sensing is to help planners make accurate short-term decisions based on demand changes that occurred hours or days ago, not what happened last year. Thus, demand sensing transcends latency issues inherent in conventional time-series forecasting methods – which assume that history will repeat itself.
Adopting demand sensing has been found to reduce near-term forecasting errors by 30–40%. It also empowers companies to rapidly adapt to sudden changes in customer needs, and facilitates the creation of data-driven supply chains.
Demand planning is a larger process that goes beyond forecasting demand. According to the Institute of Business Forecasting and Planning (IBF), demand planning leverages experience and forecasts to estimate demand for various goods at different points of the supply chain. Demand planners also participate in inventory optimization, ensuring availability of products, while always monitoring the gap between projections and actual sales.
Demand planning actually serves as the base for planning many other activities, such as shipping, warehousing, price forecasting, financial planning, and supply planning – all aimed at fulfilling demand based on data regarding the anticipated needs of customers.
Demand forecasting, however, forms the basis for all three of the above activities. But forecasting demand is a complicated process, and can be done in various ways – the two most popular ones being traditional statistical forecasting, and machine learning-based algorithms. The choice of which one to use depends largely on business type, available resources, and business objectives.
Traditional statistical forecasting
This technique has been around for years, and remains a staple choice among myriad forecasting processes. The major difference when compared with how it was done in the past is that all the calculations are now performed automatically, using modern software. For example, you can now generate time-series sales forecasts and trends in Excel.
Data sources: To forecast demand, statistical techniques use past data – which is why statistical forecasting is also often referred to as historical. Common practice dictates that for best results, it is recommended to collect sales data for at least two years.
Pros: Traditional forecasting is still a very popular approach to predicting sales. On average, demand statistical technique-based planning solutions seamlessly integrate with Excel as well as existing Enterprise Resource Planning (ERP) systems, without necessitating additional tech investment or expertise. Highly advanced statistical models can also consider seasonality and other market trends, and apply various methods to finetune results.
Cons: For statistical forecasting accuracy, stability is key as it assumes that history repeats itself (situations that took place 2–3 years ago will recur). However, this is far from true in the real world. As a result, statistical techniques are flawless in an ideal world, but often fail to predict unanticipated changes in customer preferences, or predict when market saturation will set in.
Ideal use cases: All in all, statistical forecasting (automated) offers a passable level of accuracy in the following cases:
- Mid to long-term planning
- Well-established products that witness static demand
- When predicting overall demand is more important than forecasting sales of individual stock-keeping units (SKUs)
Machine learning-based forecasting
Growing demand volatility on one hand and increased computing power on the other have given rise to the need for wider use of machine learning (ML) to increase forecast accuracy. ML also accelerates predictive analytics beyond merely estimating demand – it combines historical data with current data to generate insights into customer trends and behavior and offer real-time recommendations on how to influence both.
Data sources: While ML-based algorithms are built upon statistical models, they also leverage additional external as well as internal data sources to generate more accurate, fully data-driven demand predictions. More importantly, these algorithms can work with structured as well as unstructured data, including the following:
- Macroeconomic indicators
- Competitors’ activities
- Marketing polls
- Past financial and sales reports
- Weather forecasts
- Local news
- Social media signals
Pros: ML-based techniques apply complex mathematical algorithms to automatically capture demand signals, identify patterns, and spot complex relationships in large datasets. In addition to being able to analyze large volumes of data, these smart systems keep retraining models to adapt them to changing conditions – thus addressing the challenge of volatility that traditional methods are unable to overcome.
Thus, ML-based techniques generate far more accurate and reliable forecasts in even highly complex situations. In addition, they eliminate the need for time-consuming and tedious manual adjustments and recalibrations.
Cons: To truly leverage ML-based techniques, businesses need considerably large volumes of high-quality data. Without this, the system will not be able to establish accurate benchmarks and generate relevant predictions.
Also, in addition to the unavoidable complexity in terms of software maintenance, there is the matter of result interpretation. Despite the fact that ML-based algorithms can arrive at conclusions without human intervention, it is up to a live expert to decide which features should be fed into the model, which of them has/ have the biggest impact on output, and why the model generates a certain prediction.
Ideal use cases: Situations in which ML models definitely work better than traditional models include the following:
- New product launches
- Volatile demand patterns
- Fast-changing environment
- Short to mid-term planning
Short-term forecasts are typically more accurate
Truth be told, in today’s fast-changing market, it is far easier and more effective to predict what happens in a few days rather than a year. Starting with basic factors such as weather conditions and working up to constantly evolving customer preferences and other unstable market and macroeconomic forces, ML models are purpose-built to process reams of data for near-term, actionable insights. The shorter the time frame, the more accurate the predictions.
We understand that the switch to an ML-based model can be daunting, and that it definitely cannot be an overnight decision. That’s why we give our clients the option to start as small as possible, and work their way up based on their comfort and convenience. To get started today at zero risk, give us drop an email at sales@throughput.ai. We’ll get back to you in no time at all for a no-strings-attached, open chat on how you could benefit from ML-based demand forecasting. We promise you, you won’t look back.
