A Data Scientist’s Path to Understanding Market Simulation
Made possible by recent advances in computing power and machine learning, market simulation employs agent-based modeling, behavioral science and network science to recreate the complex dynamics and rules of how a population of people in a given market behave, influence each other and make decisions.
By Kalina Angelova, Concentric.
When I joined the Concentric team, I had no previous experience with market simulation—my data science background was in traditional econometric models and supervised machine learning. I knew that, in general, simulations are a representation of a system, its components, relationships and rules of behavior, but I wasn’t clear on the architecture of market simulations. When first introduced to the concept, I immediately thought of Monte Carlo simulations, recognizing that both Monte Carlo and market simulations account for uncertainty. I was on the right track, but there was so much more to uncover on the path to truly understanding market simulation.
Made possible by recent advances in computing power and machine learning, market simulation employs agent-based modeling, behavioral science and network science to recreate the complex dynamics and rules of how a population of people in a given market behave, influence each other and make decisions. Where traditional models assume that humans are rational actors, market simulation is unique in that it accounts for network dynamics. It captures instabilities, nonlinearity, heterogeneity, network structure and endogenous dynamics in the market and can be validated to multiple data sources at the same time. The complexity of the model is high and at the same time applicable in a diverse range of fields.
Fig. 1: Predictive Analytics vs. Market Simulation
I’ve come to value the complexity, flexibility and accuracy of market simulation, but honestly, it took me awhile to build a foundation of trust and appreciation for the method. The famous George Box once said, “Statisticians, like artists, have the bad habit of falling in love with their models.” In my experience as a data scientist, I think he’s right. And as with any solid loving, trusting relationship, it takes time to develop true appreciation for a new data science method. With market simulation, I was first skeptical, but through a little trial and error, I soon got there.
Here are three discoveries I made on the path to understanding and trusting market simulation:
Market Simulation Takes A Human-Centric Approach to Analysis
When building a market simulation, first you must develop the appropriate abstraction. The model needs to account for the unexpected, complex nature of a market in order to truly capture the dynamics.
In the real-world, people are affected by marketing activities. They experience different products and communicate those experiences with each other. Different groups of the population have different values, preferences and habits. And, people are sometimes irrational—they forget about alternatives in the market or buy by inertia. Channels and experience drive consideration and perception of the market and ultimately result in sales. In order to build an accurate market simulation, you must start by mapping of all those interconnections.
Once you’ve laid out a map of market behavior, you then adapt it to a specific industry. The market is defined by establishing the outcome you are trying to influence. Success might be measured in sales units but could also be in new subscribers, website visits, online application submissions or program enrollments. Whatever the measurable outcome, you specify all alternatives available to the consumer, and how the consumer assesses their options. Then, you plug in historical data in order to evaluate the strength of each of the causalities built into the model. This process of parameter evaluation and model refinement is called calibration.
With calibration, on an ongoing basis, values are verified by testing forecasts against observed metrics and in-market results. The model is then adjusted and tested again so forecasts and observable outcomes continue to match even as market dynamics change.
Market Simulation Makes It Easy to Run What-if Scenarios
In analytics, each model has its own advantages and disadvantages. Regression models, for example, are great at explaining causality and measuring the weight of past variables on outcomes, where other machine learning (ML) models focus more on the fit of the estimated to the historical data. In general, ML models rely heavily on historical data and so struggle with forward-looking analysis. This is where market simulation has the advantage.
With ML models, users start by entering data and establishing parameters, and then finally apply human interpretation. With market simulation, it’s the other way around. You start with human interpretation, apply the data and then adjust the parameters to fit with the real world. By capturing the market and accounting for human behavior—human perceptions, awareness and irrationality—market simulation allows users to test ideas and what-if scenarios, to measure the impact of future decisions.
Market Simulation Doesn’t Substitute Traditional Models, it Adds Additional Insight
Market simulation is a great approach to diversify your toolkit. Instead of substituting traditional models, it validates and amplifies their insight. Simulation aggregates the findings of traditional methods and offers a bigger picture and overall understanding of the system both today and in the future.
Three common applications of market simulation include demand planning, business forecasting and predictive marketing mix modeling. Automotive companies, for example, use it to optimize production capacity and new launch strategies. Meanwhile, media companies like E.W. Scripps Company use market simulation to advise brands and campaigns how to best allocate their media spend to deliver impact. Using the Concentric market simulation platform, in one year the E.W. Scripps Company created more than 2,200 models and analyses across 16 consumer data sets, achieving 84 percent forecasting accuracy when compared to in-market results. In general, organizations using market simulation find they are able to create a more unified view of their market, which enables them to reduce uncertainty about strategic investment decisions.
While market simulation is popular in physical and natural science, it’s still not widely recognized in the analytics community. The main misconception is that the forecasts it generates are random. Data scientists need to be convinced that market simulation modeling actually works. My best advice for data scientists who are curious but not yet convinced by market simulation is to try it yourself. Through trial and error, you too will see the method’s potential to complement and advance your toolkit—you may even, like George Box predicts, fall in love with it.
- Top Obstacles to Overcome when Implementing Predictive Maintenance
- How to Use Cohort Data to Analyze User Behavior
- No order left behind; no shopper left idle.