Agent-based modelling (often shortened to ABM) is an analytic approach for simulating a marketplace (or other environment) by coding individual, autonomous decision-making units (called agents) that are programmed to make decisions, interact with each other, learn over time, and/or respond to changes in the environment.
ABM is a simulation/ computational approach rather than a statistical/econometric one, but the specific rules governing how agents behave are flexible and can be driven by statistical models, Behavioural Economic "rules of thumb", or other logic.
Unlike a top-down approach – like an economic model that looks at last year's overall sales and adjusts them to estimate this year's numbers – an ABM would start from the bottom-up, simulating individual consumers (using data from surveys, client databases, and/or passive data) and turning their unique preferences and characteristics into personalised decision-making rules for each of them in the model. For example, perhaps data shows that some consumers are sensitive to peers and will purchase whichever item is currently the bestseller in the market. At certain intervals these consumers (and only this subset of consumers, if so desired) can be programmed to take note of the current bestseller in the market and select that product for purchase. In this way, one can create models that allow for dynamic market effects like snowballing or fads.
Ipsos Point Of View:
Exploring a client's market with an ABM can provide several advantages over more traditional approaches:
- Ability to integrate multiple data sources. The goal of an ABM is to build a richer, more realistic portrait of consumers. Multiple surveys, transactional databases, and/or social or behavioural data sources can be integrated under this framework of building a more dynamic, realistic set of simulated buyers.
- Ability to explore more "what if" scenarios, including interpersonal influences and effects over time. What if negative information about a product spreads through social networks, right before the holidays? What if parents love a product, but children get tired of it after several uses? How does market share depend on how many months elapse between a client and competitor product launch? By building the model at the consumer level and playing it out over time, these models offer a great deal more flexibility to simulate custom changes to the marketplace and estimate the reaction.
- Greater visibility into consumer subgroups. For each of the above scenarios, it is also possible to drill down to the agent level and see which segments or subgroups are most vulnerable to these changes.
ABMs are sometimes characterised as "toy" models. This is a reflection on how this approach is typically applied in academic literature, where a particular environment and set of decision rules are thought up by a researcher and played out in order to see whether simple rules can lead to complex or undesirable outcomes. These can lead to fascinating insights (for example, that only moderate preferences for same-race neighbours will nevertheless lead to strongly segregated neighbourhoods) but are often run in the absence of any underlying data.
However, agent-based models as executed by Ipsos are quite different. Here, the rules governing consumer agent behaviour are highly data-driven (e.g. based on responses to a choice exercise, or actual past behaviour, or responses to direct questions about behavioural influences). The environment is also data-driven, in terms of (for example) the product offerings that are made available to consumer agents throughout the year.