Why forecasting in medtech is so hard and what can be done about it

Sales forecasts play a critical role in the success of every company. An accurate and sufficiently granular forecast is critical for ongoing demand planning, brand strategy, investor communications, inventory and manufacturing, and many other core functions for any company.

However, there often aren’t enough resources dedicated to accurate and granular forecasting. In addition, forecasts are commonly based on sales leaders’ personal assessments of their pipelines rather than data. With multiple medtech clients, we have observed that sales forecasts tend to swing overly optimistic, especially with products that may be gradually declining.

Forecasting is a difficult science, especially if you need to do it consistently, over time, and at a larger scale. Moreover, this problem is often compounded by additional difficulties inherent in medtech itself.

1. It’s not feasible to create detailed-level forecasts using current tools and technologies. Medtech companies deal with thousands of SKUs with complicated interactions. The short-cut methods used to overcome that complexity fall short and don’t help the manufacturing arm of the organization plan for inventories and stock-outs.
   
2. There is too much variability in pricing and contracting. Converting units to revenue often requires understanding the specifics of individual contracts created with hospitals, IDNs, ASCs and HCOs. This is difficult (and expensive) to achieve with manual methods and limited resources. But it’s vital to make accurate forecasts, especially at the local level.
   
3. Capital equipment forecasting is hard to get right. Capital equipment and consumables are highly interdependent, so it’s critical to forecast capital equipment pipelines. But the high prices and intermittent sales make forecasting difficult and mistakes costly. One medtech client we worked with saw deviations of over 20% on capital equipment sales every month at the national level. Significantly higher deviations were also observed within countries where capital equipment sales account for a large percentage of volume.
   
4. Medtech data is often messy and not easy to work with. There is rarely a well-defined and cleaned source of data that can easily be used for forecasting purposes. Therefore, forecasts are often generated using nothing more than historical sales data.

So can we do better? Are we simply stuck with unreliable, slow, and inaccurate forecast? Of course not. Fortunately, advances in analytics, computational power and data volume in the last decade means that there are now robust approaches that can help organizations solve these challenges without resorting to direct resource scaling.

1. Leverage advanced analytics to get to the SKU and customer level. Advances in computing power means we’re no longer limited to running a few models on our computers. Cloud computing means we can easily scale up our computational capacity and run thousands of models at any granularity of interest and have the machines choose the most accurate model for us according to our criteria. By getting down to the SKU and customer level, forecasting accuracy can be improved by better understanding products and customers. For example, in one company, we saw a five to ten percent improvement in forecasting accuracy when the forecast was created at channel and segment levels.
   
2. Use image recognition and machine learning (ML) to counteract pricing and contracting variability. Contracting data are notoriously difficult to use due to the way they are typically stored as paper documents and PDFs; however, text and image recognition in today’s algorithms can help organizations read hundreds and thousands of contracts and extract key features such as pricing, volume, rebates and others, much quicker than before. Even without scanning individual contracts, projecting and predicting price levels across customer types is something ML models can help us do easier and with higher accuracy.
   
3. Apply ML and model stacking to enable better capital forecasting. Machine learning algorithms can work directly with and triangulate multiple rich data sources such as field intelligence, historical capital replacement and customer profiles to create much more accurate capital sales predictions. Similarly, model stacking allows us to tie models together so that capital equipment and consumables are no longer forecasted separately.
   
4. Evolve your data storage infrastructure to clean and unite data. Along with computational power, data storage infrastructure is migrating away from traditional databases to broader data lake environments. From such environments, data scientists can easily access data across multiple departments to improve the accuracy of forecasts even further. Even if a company is still lagging on data lake development, bringing together multiple data sources for individual modeling has become much easier and more affordable with cloud solutions such as AWS or Azure.

We live in a world that’s more complex than ever before. The amount of data being generated in medtech is also exploding exponentially with the growth of smart devices. It will become increasingly critical to be able to leverage the wealth of data that is already available to obtain a competitive advantage across all facets of business. This is just as true in forecasting as it is anywhere else. Luckily, with the advances of cloud solutions, scalable storage and more sophisticated models, it’s now within everyone’s reach to develop granular and predictive sales forecasts.