White Paper

Algorithmic Forecasting for

the Life Sciences Industry
How to bring speed, accuracy, and control to
commercial forecasting
RICK JOHNSTON, PhD, Senior Principal, Consulting Services, IQVIA
ALEXANDRA TATARU, MS, MBA, Product Lead, Consulting Services, IQVIA
Table of contents
Executive overview 3

Forecasters are struggling with manual forecast tools and processes 4

What is algorithmic forecasting, and why does everyone want it? 5

What’s in it for forecasters? 6

Case Study: Excel Falls Short Predicting Biosimilar Trends 7

Algorithmic forecasting: reality vs. hype 7

It’s all about the data: an ecosystem of data handling tools 8

Automated forecast algorithms: the self-service analytics revolution 9

Collaboration is key: a real-time forecast workflow 11

Where do I begin? The algorithm journey 12

Conclusion 13

About the authors 14

Executive overview
Imagine being able to integrate thousands of pieces of market and product
data into your commercial forecast – in just minutes. Now imagine being able to
update that forecast in real-time, as often as you like, with the most current data
available. With manual commercial forecasting this would be impossible. But for
companies that have implemented algorithmic forecasting, it’s the new normal.

Algorithmic forecasting uses real-time data feeds, commercial forecasting teams have been slower to
automated algorithms including machine learning, and join the ‘algorithm revolution’, the growing pressure
an updated forecast workflow to predict commercial to analyze more data faster as well as the potential
performance. Its premise is simple: automate forecast for greater accuracy is pushing forecast teams to look
updates as new data becomes available, and reduce for better and more streamlined approaches to use
human manual labor in the forecast process. This brings automation and machine learning.
speed, accuracy, and deeper insights to commercial
This whitepaper offers forecasting teams a framework
forecasts and frees up forecasters to spend their time
and roadmap for adopting algorithmic forecasting.
on strategic tasks like identifying exactly what questions
Whether they have already begun their digital transition,
need to be answered, interpreting outcomes, and
or are still using spreadsheets to manage data, this
informing product strategy.
paper provides practical steps and best practices for
This type of automation and analytics is already deploying the technology, generating early wins, and
being used across the life sciences industry to speed convincing stakeholders of the strategic benefits of
insights and support informed decision-making. While embracing this real-time forecasting approach.

iqvia.com | 3
Forecasters are struggling Global and local teams often struggle to strike a balance
with regards to how tailored should the forecast
with manual forecast tools structure be to each market, who owns the forecast, and
and processes to what extent expert opinions should be allowed to shift
the forecast. Global and local teams typically maintain
In the traditional forecast process, data is manually
multiple versions with no single source of numbers to
gathered by forecasters and entered into a forecast
drive decision making.
environment, most often an Excel spreadsheet. During
this process, forecasters may find themselves spending CONSISTENCY
days, weeks or months researching, exchanging messages
with other team members, and tweaking numbers to see “There’s not a lot of consistency in the way
the impact on results. Those numbers then pass through the team does forecasting – we have some
a sequence of review by other groups within the company home-grown models, some from vendors…
– management teams, supply chain, and potentially others We don’t really have an end-to-end
– each of which may request additional iterations based
forecasting process.”
on their view of what they expect.
Small and medium sized companies often struggle
There are a number of types of issues that forecasters
with setting up an infrastructure that enables them to
have expressed with this traditional process:
maintain consistency across products in their portfolio.
Especially when there’s a lot of business development
ANALYSIS TIME
and early stage forecasting, leadership may find that the
“It takes weeks to get results when I ask comparisons are not “apples to apples”.

for an alternative scenario, and I don’t

A DELUGE OF DATA
know why.”
“We are under-utilizing existing healthcare
When so much of the data collection and entry is manual,
data we purchase. For forecasting purposes,
or potentially involves interactions between teams in
different countries or continents, turnaround times can
we only review our sales data once per year,
be long. Leadership gets frustrated at the slow response, but it is updated every month.”
and forecasters get frustrated when seemingly small
The supply of healthcare data has grown faster than
requests add up to a large load on their time.
forecasting teams’ analysis capabilities, and teams are
GLOBAL VS. LOCAL looking for ways to better integrate these sources of
information. Anchoring forecasts in real-time data also
“Coordinating forecasting activities reduces opportunities for forecast bias – where well-
between global and local teams has been intentioned forecasters or leadership may skew the

a big challenge. Local affiliates currently forecast based on personal expectations.

drive the forecast, and global has poor

visibility into how assumptions drive the
final numbers.”

4 | Algorithmic Forecasting for the Life Sciences Industry

What is algorithmic forecasting, update the forecast in real-time as new data becomes
available. A combination of human users and quality
and why does everyone want it? control algorithms validate the auto-generated forecast
to guard against errors in the process.
Algorithmic forecasting provides a way to address many
of these challenges.
A key element of algorithmic forecasting is that it is not
human-mediated.
In algorithmic forecasting, a combination of analytics
technology and machine learning algorithms replaces
Forecasters set up and supervise the process, but they
most of the manual labor, data chasing and cut-and-
are not actively involved in scouring the data and making
paste activities common in a traditional forecast
predictions. That is all up to the algorithm(s), which
workflow.
search the data for meaningful trends, and generate
predictions in a fraction of the time it takes to conduct a
Sources of relevant data are directly connected to the
human review.
forecast engine, allowing the applied algorithms to

HUMAN-MEDIATED FORECASTS (TODAY)

Data Import “Algorithm” “Validation” Final Forecast

Manually collected and Forecaster ‘runs the A user manually checks Built and maintained
copy-pasted into the forecast numbers’ and tweaks items the numbers by a person

Connected Data Automated Algorithms Validation Evergreen Forecast

Updated automatically when Data is adjusted and trended QC algorithms along with Automatically remains
source data is updated then fed into the forecast users sense-check forecast up-to-date with latest
predictions data inputs
ALGORITHMIC FORECASTS (FUTURE)

Setup and Supervision

Forecaster supervises the process

Whenever new data is added, or updated, the system for a complex submission process. In this environment,
automatically re-calculates the forecast. With robust the forecast remains ‘evergreen’ – universally and
version-control built in, forecasters can track and continually relevant, rather than updated only once each
compare forecast results, and also face a reduced need month or quarter.

“Evergreening” – where the most accurate forecast is always available,

rather than updated once each quarter – would be revolutionary
for commercial pharma forecasts. And algorithmic forecasting is the
way to do it.”
RICK JOHNSTON, PhD, SENIOR PRINCIPAL

iqvia.com | 5
What’s in it for forecasters? are interested in algorithmic forecasting, according to a
recent survey conducted by IQVIA.
With these manual steps becoming more automated,
And because the algorithm has no stake in the output,
the forecaster is elevated to a more strategic role.
it reduces or removes human bias, which was also
Because the forecast inputs are automatically updated,
ranked among the top three reasons why forecasters are
forecasters can instead focus on exploring trends in
interested in incorporating algorithmic approaches in
the data, conducting what-if or uncertainty analysis,
their forecasts. This real-time capability transforms the
identifying key drivers or indicators and developing
forecast from a quarterly static snapshot to a constantly
strategic insights.
evolving review of trends as they are occurring.
These benefits ultimately lead to more accurate
forecasts, which is the leading reason why forecasters

Why are your teams interested in algorithmic forecasting?

Increasing the accuracy of predictions

Automating forecast inputs to allow

forecasters to focus more on analysis

Reducing human bias

Reconsidering the entire approach

to the forecasting process

Looking to improve usage of

the data we purchase

0 10 20 30 40 50 60 70 80

Percentage of respondents choosing this option

407 respondents, survey date Feb 2021

Because the forecast inputs are automatically updated, forecasters can

instead focus on exploring trends in the data, conducting what-if or
uncertainty analysis, identifying key drivers or indicators and developing
strategic insights.

6 | Algorithmic Forecasting for the Life Sciences Industry

 Case Study: Excel Falls Short Predicting Biosimilar Trends
A large pharmaceutical company wanted to track programmed to automatically segment the providers
the erosion of sales volumes related to the release into ‘similar groups,’ and to aggregate their sales data
of several biosimilar products in the previous three and remove outliers. The result was a model split into
years. Sales reps were seeing unusual trends in the 2000 segments of hospitals with similar patterns.
market, characterized by some hospitals continuing The depth and detail provided through this forecast
to prescribe the innovator product while others gave them significantly more predictive results in a
shifted to the biosimilar. fraction of the time.

The company’s forecast engine was an Excel THE RESULT

spreadsheet that could only project trends in 40 A 50x increase in the number of segments made
market segments. It wasn’t detailed enough to it possible for an incredibly granular view of the
“tease out” patterns of behavior that were occurring. behavior of similar hospital groups. With aggregated
As a result, the company was unable to determine data illustrating overall patterns by provider, site
which of its marketing and sales strategies were of care, individual hospital network, and payer
having the desired impact, and what was causing trends, forecasters were able to build more accurate
some customers to switch. forecasts. Every time new data was added to the
data lake, the forecast automatically updated to
THE SOLUTION
re-calculate the impact, with no human involvement
In response, the company deployed IQVIA’s Forecast
required.
Horizon to generate a more sophisticated, real-time
market forecast. With its algorithmic forecasting The company used the forecast to make significant
platform, they were able to analyze data from changes to its sales and marketing strategy, providing
tens of thousands of individual providers and actionable tips to teams that reduced overall attrition
match it to payer-level data, thereby determining for the brand, allowing it to retain a significant market
what percentage of the innovator molecule was share despite strong competition.
prescribed versus the competitors. Algorithms were

Algorithmic forecasting: industry to meet the needs of pharma. But pharma

commercial forecasts are different from other markets.
reality vs. hype The average pharma forecast we encounter has more
than 80 variables being input and calculated, while the
While life sciences industry vendors have been touting
typical consumer goods forecast has less than 10. The
the “wonders” of automated solutions for years,
main reason for this is the need to model the complex
algorithmic forecasting has remained an elusive target
patient dynamics of most modern medicines: periods of
for pharmaceutical companies. That has caused some
therapy, patients dropping off therapy for a variety of
pharmaceutical companies to grow skeptical of the hype.
reasons at different times during their therapy, complex
Part of the problem is that a lot of forecasting dosing patterns, treatment line changes, flow of patients
applications on the market today are built by companies between lines of therapy, and others.
that try to adapt a forecasting solution from another

iqvia.com | 7
The forecasting tool also has to be able to account for smoothing tasks. These are commonly considered ‘pre-
local regulations, emerging healthcare events, and other processing’ steps that are performed before passing
issues that are unique to this industry. data to a forecasting platform. Examples of algorithms
that might be applied include isolation forests,
Practical algorithmic forecasting solutions today are built
minimum covariance determinate, local outlier factor
around three core components:
methods (outliers) and Friedman super smoothing.

1. An ecosystem of data handling tools that can pull

• Save high-fidelity data and allow aggregations:
real-time data feeds from a variety of sources and
While time series data is often saved at a daily or
employ data management methods to prepare and
weekly granularity, most forecasts operate at a
combine the data for reliable, robust analysis. In many
monthly or yearly granularity, so data is usually saved
cases these re-use data infrastructure developed for
in the more detailed form. Additionally, the ability to
other analytics and insights tasks within the business.
aggregate data across segments, such as individual

2. Automated algorithms using self-service analytics hospital data across a zip code, city, state or country,

that have been trained using vast global healthcare is required. Pivot table type functionality here can

databases to identify trends of relevance and provide help to aggregate the data to a format that provides a

insights that drive decision-making. Critical to these consistent segmentation along with time series data

algorithms is the ability for forecasters to make edits at the right level.

themselves using self-service analytics tools.

• Experiment with different approaches: Data should

3. Real-time forecast workflows that allow teams to be available in a form that allows users to quickly see

conduct frequent updates, ask follow-up questions, and perform custom queries with it. These tools are

and model many detailed scenarios to enhance typically workbenches that enable data science teams

understanding of key trends and monitor changes to use open source components like Python and R to

over time. do complex data transformations, and then save those

scripts to run each time new data is available.
This combination of elements makes it possible for
• Build automation steps for data updates: To support
forecasters to enhance the quality, detail and reliability
the algorithmic forecast, data will be coming in at
of their forecasts, and to make forecasting a more active
regular intervals. After data has first been identified
part of the strategic decision-making process.
and transformed, the infrastructure needs to be able
1. IT’S ALL ABOUT THE DATA: to perform a robust scheduled process to push the
AN ECOSYSTEM OF DATA HANDLING TOOLS analytics to the forecast.
The first enabling step for teams to implement
Most pharma companies today will already have the
algorithmic forecasting is to build a data ‘landscape’
capabilities within their data analytics teams to handle
that supports real-time data feeds in a format that can
many of these tasks. A rapidly evolving ecosystem of
be understood by a forecasting platform. These data
tools is also increasingly able to automate many of these
handling tools are not just a database or data lake, but
data handling steps.
include capabilities to:

• Join, filter, and smooth data: Data handling tools

need to be able to join data from disparate data
sets, remove outliers and errors, and perform data

8 | Algorithmic Forecasting for the Life Sciences Industry

2. AUTOMATED FORECAST ALGORITHMS: speed of their forecasts. The evolution of algorithmic
THE SELF-SERVICE ANALYTICS REVOLUTION forecasting can provide far more sophisticated solutions.
While Excel has been the backbone of forecast
IQVIA has managed forecast processes and built many
technologies for decades, it is generally insufficient by
of our own Excel forecast platforms over the years. We
itself for algorithmic forecasting. There are issues with
have learned that to enable algorithmic forecasting, a
it having limited flexibility, challenging version control,
new type of forecast enterprise architecture is required.
and sluggish performance on large models that creates
headaches for forecasters and reduces the agility and

“Insights and Analytics” function

Robust method
to retrieve data
Ingestion

Extract,
transform, Wrangling Forecast Engine
load data Integrate Integrates multiple
metrics into a
coherent forecast

Computation
Feature detection,
machine learning

A key part of that technology is to think about the way There are a whole host of different algorithms that can
that open-source tools like Python and R – as well as the be used to process forecast data depending on the type
ecosystem of data handling tools around them – can be and detail of the data, objective, and capabilities of the
integrated into the forecast process. IQVIA’s approach algorithm. We list a few categories of algorithms below,
has been to build a platform that allows users to use but there are also many different publicly available
their own insights and analytics tools, but to allow algorithms that can be tuned to your data set.
the ‘hooks’ to a forecast engine that brings all those
computations together into one user interface.

IQVIA’s approach has been to build a platform that allows users to use
their own insights and analytics tools, but to allow the ‘hooks’ to
a forecast engine that brings all those computations together into one
user interface.

iqvia.com | 9
 THREE TYPES OF ALGORITHMS TO CONSIDER

Time Series Predictions Clustering and Partitioning Forecast Quality and Tracking
Machine Learning is creating more Important for identification of Methods to detect when the
accurate methods for trending analogues and other ‘similarities algorithm is failing and a person should
historical data within groups’ intervene to see what’s happening

Example: Prophet, XGBoost Example: K Means, KNN Example: Trigg’s Tracking Signal

There are three types of algorithms that can deliver early wins for life sciences forecasts:

Time Series Prediction Clustering and Partitioning Forecast Quality and Tracking
algorithms. In these cases, algorithms. These are suitable algorithms. Once an automated
historical data is used to predict for patient identification, market algorithm has been established, it
future performance. Algorithms segmentation and analogue is critical to ensure that the outputs
like ARIMA and Holt-Winters are selection. In the case of patient are checked each time new data
rapidly being replaced by more identification, these algorithms can is received. These quality checks
sophisticated methods that use be used to identify undiagnosed allow forecasters to know if their
machine learning as well as more patients based on similar patient forecast algorithm is still producing
categoric variables to provide diagnoses or patient journeys. a good result or needs manual
better future predictions. They can also be used for analogue intervention.
selection, allowing similar products
to be found in the huge universe of
available marketed products based
on a set of features.

Sources:
https://towardsdatascience.com/time-series-forecasting-predicting-stock-prices-using-facebooks-prophet-model-9ee1657132b5
https://aws.amazon.com/blogs/machine-learning/k-means-clustering-with-amazon-sagemaker/
https://pdfs.semanticscholar.org/c4e6/6029e5173d25a28be7fc9b811d96c675b9e3.pdf

Our focus so far has been on the infrastructure and users commit to changing their workflows and processes
technology needed to enable algorithmic forecasting. to make it part of how they operate. That’s a change
However, even the best technology can’t work unless management issue, not a tech issue.

10 | Algorithmic Forecasting for the Life Sciences Industry

3. COLLABORATION IS KEY: The shift to algorithmic forecasting requires re-
A REAL-TIME FORECAST WORKFLOW imagining these processes for the age of multi-user
The primary challenge for many organizations editing and real-time data. Commercial forecasters need
attempting to embrace algorithmic forecasting is siloed to learn ‘the basics’ of these analytics tools and how they
responsibilities. Today’s forecast processes – the whole might be used, while insights and analytics teams need
system of submissions and reviews – makes it difficult to learn to apply their toolset to practical real-world
to work cross-functionally in the way that is necessary forecasting problems. That includes determining how
to implement algorithmic forecasting. In the same vein, data can automatically update a forecast without explicit
there is a lot of momentum behind those processes as approvals, and how it can automatically check the errors
they cross many groups in the organization. and flag any issues which require human intervention.

The forecast is at the center of a collaborative effort In the collaborative process

• The forecast is always live and can

Brand Global
Forecaster Forecasting be updated by any user in
real-time

Orchestrate Review and • The brand forecaster

Forecast Add Data “orchestrates” the changes and
Build versions to ensure accuracy
Real-Time Build
Data Links Algorithms • The mechanics of forecasting are
Data Analytics
Stewards mostly/entirely automated
Collaborative • The latest numbers are always
Forecast
Review available, as well as a history of
Data Inputs Questions
and Provide and Scenarios changes
Assumptions
• No formal ‘submission’ is needed
Other Groups Senior • Teams have high visibility into the
(marketing, supply chain, Leadership forecast framework, assumptions
finance)
and processes

All of this technology already exists today, but the key The results of this type of process speak for themselves
is to make the business process changes required to – eliminating differences between the ‘official’ and
make this collaborative environment a reality. That the ‘latest’ versions of the forecast, enabling all team
means doing away with cumbersome submission members to access one source of truth which is always
technologies, versioning systems, review processes and kept up to date with the latest numbers, and removing
other paradigms designed for the age of Excel. Instead, sources of bias introduced by different groups during the
a forecaster becomes the conductor of a team of people hand-off process.
all providing forecast data and analytics in real-time,
‘orchestrating’ the entire process from end-to-end.

iqvia.com | 11
Where do I begin? approach where parts of the forecast funnel are
transferred to algorithmic-type approaches, while
The algorithm journey others are still manually forecast. This approach focuses
algorithmic forecasting on areas where it delivers the
Algorithmic forecasting will not be a simple process for
most value.
most companies. IQVIA recommends an incremental

HOW DO WE DETERMINE IF A FORECAST INPUT SHOULD BE ALGORITHMIC OR MANUAL?

Better for Better for

manual forecasting algorithmic forecasting

Real-world data sources available

Data in offline systems
for the metric
(e.g. academic papers)
Data changes significantly
Data changes infrequently over years
from month-to-month
A meaningful change to input
Small changes to input have
doesn’t have a large impact
a big impact on key metrics
on key metrics
(e.g., volume / revenue)

Algorithmic approaches are favored for inputs where: Common examples include market share, total
volume, diagnosed or treated patients.
• Detailed real-world data is available and regularly
updated Prioritization of those metrics with greatest impact
can be informed with a tornado-style analysis that
• Data is changing significantly from month-
looks at potential ranges for those variables and the
to-month
effect on net revenue.
• A small change in that input would cause a
significant change to forecast outputs, such as
net revenue or total volume

12 | Algorithmic Forecasting for the Life Sciences Industry

Many forecasters ease into algorithmic forecasting Users can specify complex patient dynamics such as
by taking a ‘manual-lite’ approach, where algorithms persistence (time on therapy) and flow between lines
suggest a solution, but confirmation is required from of therapy, reflect the patient journey across different
the forecaster to actually apply the algorithm in each demographics, and compare the performance of
case. For example, the algorithm may suggest a ‘naïve’ different products across treatment categories. These
forward-looking trend to the user – lowering the chance automated insights enhance the depth, accuracy, and
of bias in their estimations – but it would still be up to precision of forecasts, while minimizing time spent on
the forecaster to accept the recommendation. Analytics- manual updates.
assisted inputs like these still deliver many of the benefits
Both platforms are helping life sciences companies
of algorithmic forecasting, without total automation.
across the industry navigate forecasting challenges and
The complexities of forecasting in the healthcare space, allowing them to evolve their approach to real-time,
as well as the importance of accuracy means pharma evergreen forecasts.
companies may struggle to find forecasting technology
capable of meeting their needs. That’s exactly what
IQVIA’s Forecast Horizon and Pipeline Architect
Conclusion
platforms were built to do. Commercial forecasters are looking to embrace
algorithmic forecasting as a method to achieve better,
These two solutions were built by experts in data
faster, and more accurate results.
analytics, global healthcare regulations, and digital
technology to ensure the platforms could meet the exact However, many life sciences companies are uncertain
needs of pharmaceutical industry stakeholders across about where to begin. The good news is that it isn’t as
the globe. hard as it sounds.

IQVIA’s Pipeline Architect1 platform, designed for early With this framework, forecasters can begin discussions
stage products, uses algorithmic forecasting to pull to put together their data infrastructure, determine what
information on hundreds of trials from IQVIA’s global algorithms suit their objectives, and begin discussing the
databases and combine that with public data to predict potential benefits of a real-time forecasting process. And
an asset’s development costs and timelines, as well as to even if they are not ready to jump in fully, there are ways
inform high-level development timing and commercial to benefit from algorithmic forecasting, like automating
expectations. These insights lead to better forecast a smaller number of inputs or requiring manual
development and more accurate timeline predictions. approval, which can be put in place today.

IQVIA’s Forecast Horizon2 , designed for later stage

or inline products, is a commercial forecasting tool
which can be used to build an algorithmic forecast.
The platform can pull in data and from clinical and
commercial data resources and apply out-of-the box
or custom-built algorithms to inform forecast inputs.

1 https://www.iqvia.com/solutions/research-and-development/consulting/pipeline-architect
2 https://www.iqvia.com/solutions/commercialization/brand-strategy-and-management/forecast-horizon

iqvia.com | 13
About the authors
RICK JOHNSTON, PhD ALEXANDRA TATARU, MS, MBA
Senior Principal, Product Lead,
Consulting Services, IQVIA Consulting Services, IQVIA

Dr. Johnston is a Senior Principal and Alexandra is a Product Lead for

Software Lead within IQVIA’s Consulting Organization. IQVIA’s Consulting Services team. She focuses on the
He is an accomplished technical leader with more than blend of technical and business concepts within the
15 years’ experience in building software that solves big life sciences space, with over 9 years of experience
data, machine learning and advanced analytics problems in healthcare across a mix of product management,
in pharma. Software he built is used by more than 90% management consulting, and engineering. Alexandra
of the world’s 25 largest pharmaceutical companies. holds a BS and MS from Stanford University in
Dr. Johnston has a PhD in operations research and Mechanical Engineering as well as an MBA from
bioinformatics from UC Berkeley. Columbia University.​

14 | Algorithmic Forecasting for the Life Sciences Industry

iqvia.com | 15
 CONTACT US
[email protected]
iqvia.com

© 2021. All rights reserved. IQVIA® is a registered trademark of IQVIA Inc.

in the United States, the European Union, and various other countries.
05.2021.CAS