June 29, 2020
Forecasting: The conundrum of accuracy and scale
Forecasting. The oldest solved problem which continues to challenge businesses from times immemorial.
The ability to consistently and accurately predict any future state of business (be it for sales, revenue, customers visiting your stores, or which product they buy or how much they will spend when they visit) is a highly rated business imperative. The more transactional the business, the more important it is to be able to get a handle on what the near and mid-term future is going to be for the chosen KPI we track.
Often, the inability to get a good forecast leads to significant business impact, revenue is one of them. The goal of a good forecast is to get you enough confidence to make sound business decisions. Forecasts are generally not the end – but a means to an end.
When we evaluate good forecasting frameworks – Accuracy and Scalability stand out as the two pivots which are probably the most important – but often one of them is sacrificed due to the complexity they together bring into the process. That is ironic – as the relevance of a forecast increases when it gets more granular, but at the same time, the effort needed to build accuracy of the forecast increases exponentially with granularity!! Most people focus on choosing accuracy over granularity – thus sacrificing business insights and control at the ground level.
Let’s take an example to illustrate this point.
Say you are a Retailing Business – and have 1000 products that are sold across 1000 outlets. The most accurate form of forecasting, in this case, is when you can predict demand for each product by each outlet. This is so because each product can have different demand frameworks at each outlet owing to many parameters, including external influencers of demand like location, local weather, buying behaviors, etc.
This essentially means that you would need to build 1000×1000= 1 million forecasting models. Now, this is a tall order for even the biggest of companies!
There is another dimension of scale in forecasting – and that makes the complexity 3 dimensional.
In the above example of 1000 products sold across 1000 outlets – add the complexity of Daily Forecasting Frequency. Now that becomes a very difficult problem to address – you will need to build and run 1000x1000x30 = 30 million forecasting models on a monthly basis.
That is building 1 million fairly accurate models every day!!
Now layer this with 100’s of internal data points, coupled with many external demand influencers (Weather, Rainfall, Pricing, Macro Economic factors etc) – and you have a massive scale problem at hand. Add to this the complexity of training machine learning frameworks to build forecasting models for your specific use case which are reliable and accurate.
Most companies would never go this granular to the extent of outlet level. Hence companies consolidate the forecast at a higher level of hierarchy (Regional, Warehouse, Distributor etc) to build a lesser number of, but more accurate forecasting models. Even at a much lesser scale, they would probably do this manually – which means a big drain on time and resources as well.
Scale becomes the casualty here. Or accuracy gets compromised to get higher granularity.
Great forecasting frameworks address the problem of scale AND accuracy. If you solve for both – you would not only score high on reliability benchmarks – but also save significantly on time and costs.
FORECAST² allows organizations to build multivariate forecasting models at scale in a matter of a few clicks – generating 100’s of models within minutes and hours. Focused on non-technical users – FORECAST² allows business users to build granular forecasts without engaging in complexities of data frameworks, ML models, parameter correlation, etc. In a few clicks – business users can upload data in Excel sheets, choose forecast predictors – and start generating accurate forecasts quickly.
With the ease and flexibility of deployment – an organization could get started almost immediately on building complex forecasting models – at a TCO which is unmatched in the current tools available. Simple data requirements mean that the prep work required to build accurate forecasting frameworks is almost minimal, and business users can do it on their own. It also allows deep levels of Scenario Planning and What-If Analysis – leading to better and more agile decision making.
So, if you are a business selling 1000’s of products across 1000’s of stores (like in our example above) – you could start forecasting your daily demand by product across stores very quickly. As you append new data – the self-learning models align even more closely to the KPI’s which you are trying to achieve with forecasting.
Accuracy is the bedrock of forecasting. Scale brings agility in business planning which is amplified by granular forecasting.