Using predictive analytics and data science to forecast drugs price in the USA.

Disclaimer: Any views or opinions represented in this presentation are solely attributable to the authors and are not indicative of the perspectives or opinions of Nephron Research LLC or Nephron Research analysts

Introduction:

This project is a collaboration between Columbia University and Nephron Research LLC, a healthcare equity research provider. The goal of the project was to analyze pricing and volume dynamics in branded drugs, in order to provide Nephron and their clients with actionable insights and predictions for the branded drug market. Could we find patterns among the various drug groups and build a predictive model to anticipate drug price changes?

Data and Methods:

Our data consisted of historical pricing (WAC price) and volume data for thousands of branded prescription pharmaceutical drugs starting from 2014, along with other categorical information about the drugs, such as their classification. Our approach was first to perform exploratory data analysis, in order to understand how drug price, volume, and other categorical data interact. We then incorporated these insights into a predictive model for future drug price trends. As our analysis progressed, we uncovered new trends and patterns in the underlying drug pricing data that allowed us to increase the accuracy of our predictive models.

In our initial analysis, we found strong seasonality for historical drug price changes. The below chart displays the average month-to-month, price percent change for all drugs in our dataset.

Source: Nephron Research Data

This visualization, along with other analyses, revealed that the vast majority of drug price changes occur annually at the beginning of the year. While some drug prices decreased, the industry is dominated by drug price increases that follow a consistent year-to-year pattern. In the last couple of years, this pattern has held, but the rate of increase has slowed.

Prophet Model:

Following this analysis, we focused our attention on predicting when and by how much drug prices would change in the future. Ultimately, our goal was to make predictions at the drug class level, focusing first on individual drugs to calibrate our approach and better understand drug price behavior. We quickly determined that Prophet, a time series forecasting model developed by Facebook, gave us the best results. Prophet is based on an additive model where non-linear trends are fit with yearly, weekly, and daily seasonality, which makes it works well with time series with strong seasonal effects, such as drug prices.

Here is an example of our Prophet model for a specific drug, Synthroid. The dots are the actual observed monthly prices, and the line is the model’s prediction. Dark blue represents where we trained the data, while light blue represents the test set and our prediction

Source: Nephron Research Data

Synthroid is representative of the Prophet model’s excellent performance. We built upon this success to make predictions at the drug class level, using the sales-weighted price for each class in order to accurately incorporate each drug’s relative importance within a drug class. Below is an example of the Prophet model’s performance for Vaccine drugs.

Source: Nephron Research Data

In order to achieve these results, we did have to make some tough calls. First, we excluded drugs that approached patent expiration, after uncovering abnormal price behavior as drugs approached expiration. Theoretically, we could have added patent expirations date as a regressor to the model, but we uncovered that this behavior varied greatly by drug class, resulting in inconsistent predictions. Second, we excluded drugs with incomplete historical data, as many drugs change their indication and as a result their ID number. With each exclusion, we created a data universe that was less representative of the real world; however, we believe we struck the right balance by understanding the implications of our choices and investigating the characteristics of the data we planned to remove.

Our Predictions

With a well-performing model built, we focussed on the following drug classes:

• Allergy, Systemic, & Nasal
• Dermatologics
• Gastrointestinal (GI) Products
• Immunology
• Multiple Sclerosis
• Nervous System Disorders
• Oncologics
• Ophthalmology, General
• Other Central Nervous System (CNS)
• Other Haematology
• Vaccines (Pure, Comb, Other)

Most classes we investigated followed overall drug market trends; however, the Other CNS and Vaccines classes were outliers.

Sales-Weighted Average WAC Price Month-Over-Month Percent Change

Source: Nephron Research Data

Unlike the rest of the drug market, Other CNS saw sales-weighted prices decrease, and as shown below, we predict this downward trend to continue into 2021 and beyond.

Sales-Weighted Average WAC Price for the Other CNS Drug Class

Source: Nephron Research Data

When compared to the Multiple Sclerosis and Nervous System Disorders classes, it is clear that Other CNS has been much more volatile, and our Prophet model anticipates this to continue into the future. Perhaps there is a relationship between price volatility and decreasing prices and an area worth exploring.

Sales-Weighted Average WAC Price Month-Over Month Percent Change for Other CNS, Multiple Sclerosis, and Nervous System Disorders Drug Classes

Source: Nephron Research Data

Vaccines, on the other hand, see large price spikes relative to the market in the early fall. Of the drug classes we assessed, Vaccines are the only class to display seasonal price variations, likely related to flu vaccination in the fall. Our Prophet model predicts these seasonal fluctuations in sales-weighted average WAC price will continue in the future.

Sales-Weighted Average WAC Price for Vaccines (Pure, Comb, Other) Drug Class

Source: Nephron Research Data

A final comparison we found interesting was oncology and immunology; two classes affecting the drug portfolios of many pharmaceutical manufacturers. While both are projected to see price increases in the future, our models show that we anticipate immunology will see more rapid, consistent price increases, while oncologics will increase at a slower, choppier pace.

Source: Nephron Research Data

Conclusion

Prophet is a great tool that exceeded our expectations for predicting drug prices. We look forward to seeing how well our model performs in the coming year, especially to see if COVID-19 has an impact on our prediction’s performance. In addition, we hope that these trends will spark investigation by investors to uncover the industry dynamics driving these relationships.