An industry-specific perspective on building a forecast
I remember asking my colleagues/seniors one question when I started working on ARIMA Forecasting for a drug volume and other members of the team working on predicting the next three months’ drug volume using the Random Forest model. The question was," Is there any difference between Prediction and forecasting?" And the answer I used to get was ‘Yes,’ but I was never satisfied. During their discussions, I realized that there was a perception among them that both are very different. But over time, when I worked on a 6-month forecasting project, I realized that there is a fine line and forecasting is just one of the most common applications of prediction modeling. So, what exactly is forecasting? And how it fits in the area of predictive analytics?
What is a Forecast?
Forecasting for an average person is the best judgment (simply put, Prediction) about the unseen future (main keyword) based on present and past scenarios, while Prediction is simply estimating outcome based on unseen data. So, to sum up, forecasting is a subset of Prediction with temporal information, i.e., time-series data. A lot may people are working on forecasting accurately, but this is one of the most challenging tasks in the area of Machine Learning because of the uncertainty attached with the future. Many have attempted and failed because of either overestimating the actual numbers or achieving far more than expected. Hence, it is an ongoing improvement process and will see many new methods emerging in the market.
What is the purpose of the forecast?
To delve further into pharma forecasting, the first important question is the purpose of forecasting for any pharma company? It could have varied uses- such as revenue planning, production planning, resource allocation, project prioritization, partnering decisions, compensation plans, lobbying efforts, and so forth. This makes forecasting more challenging as it needs to cater to diverse stakeholders. The other complexity with forecast utilization is whether it is used unidirectionally (fed in some other function) or bidirectionally (used to quantify effects/contribution of other market events in the forecast, called "forecast function"). So, it is crucial to understand the need of recipients of forecasts to select the best method.
Image: Forecast links to other functional areas:
What is there in the future?
By the end of project completion, I understood that the next most significant aspect of forecasting is to record input scenarios and assumptions considered by stakeholders in driving the forecast, as these are used to guide resource allocation decisions. The forecaster may not be capable of gathering all areas impacting the brand where other functional teams come into play. Please note that this is an ongoing process, adjusting forecasts time-to-time based on changing resource allocation.
Forecasting is not a number!
On the other end, it is required to emphasize that forecasting is generally not a single number. But it is common in the industry to provide a range of forecasts with the probability of occurrence with confidence incorporating variations in future uncertainty attached. This is termed Probabilistic forecasting.
Build your very own first forecast model
Now, as a forecaster, I would move right onto the next section that is how to create a forecast.
A. Defining the need for the forecast (already discussed in the above section)
B. Gauging the final audience of the forecast, i.e., business/process focussed or technically heavy?
C. Market research on dynamics of historical brand sales & events(new/in-market)
D. Selecting the appropriate methodology and including relevant market resources in the model.
E. Finalizing on the best future assumptions and Analysing the forecast results
F. Presenting to the end-user, i.e., all stakeholders.
G. Adding feedback loop in the forecast models for future adjustments.
Image: The process of forecasting:
B. Why gauging the target audience is required?
In each phase of my project, I felt that though initially, the client wanted to leverage technically heavy advanced predictive analytics like tree-based models for forecasting, they realized later that it might not be transparent enough to be circulated to a broader audience within an organization due to aura of ‘black box’ attached with complex machine learning models. Moreover, there was a realization that it is hard for end-users to make simple changes to such a model to adjust the forecast frequently. Consequently, we built a simulator over our machine learning models that allowed the client to tweak projections based on different scenarios.
So, I would say the real responsibility lies on the forecaster to provide such forecast methods that meet the audience’s needs.
C. It is most essential to understand how the brand is placed in the market, whether it is a new product or is a mature one. If it is a new product, predictive modeling might not be feasible due to the paucity of data, and market research will be preferred (although a similar placed old analogy can be considered). In my case, the drug was a mature blockbuster, one means in-market with enough historical data (5 years) to proceed with. Hence, I spent significant time understanding past monthly time series sales to find trends, cycles, deviations, growth month over month.
Now, the next most important question is how to choose that "right" forecast method?
- During initial discussions in our project, the initial conversation was on the time horizon of the forecast. This is very important to know beforehand to narrow down the best method and modeling structure. As I had to do a one-year model forecast, I chose to roll up monthly level forecasts to increase accuracy.
- The next question was the forecast level; for me, it is national and district & payer level (which we called sub-national) forecasts as these drive business development for the client.
Conclusion
So far, we have covered what is forecast, the need/purpose of it, challenges with forecasting in general, and later moved to build it from scratch from a forecaster’s point of view. The next article will cover the most critical part i.e. development of the model and presentation of final short-term forecasts.
I will continue with detailed model codes/ flowchart in the next post.
Note: Some references/pictures taken from Arthur G Cook’s Book: Forecasting for the Pharmaceutical Industry
