Agent-Based Models in the Peanut Butter Industry

Why Peanut Butter?

Peanut butter is a commodity market with brand differentiation, price-sensitive consumers, volatile input costs, and a supply chain that spans farmers, crushers, manufacturers, and retailers. It is also a useful testing ground for agent-based models. The dynamics are simple enough to reason about clearly, but rich enough to produce non-trivial emergent behavior. If you can model peanut butter, you can model most fast-moving consumer goods markets.

What the Model Looks Like

A typical ABM for this market has three layers of agents.

Consumers make purchase decisions at each timestep. Each consumer has a brand preference score, a price sensitivity, and a stock of household inventory. When inventory drops below a threshold, they visit a retailer. Their choice among competing brands depends on price, shelf visibility, and a loyalty parameter that decays slowly if their preferred brand is out of stock or repriced.

Retailers manage shelf allocation and reorder decisions. They observe sales velocity and submit orders to manufacturers on a lag. Their reorder rules are simple — typically a base-stock policy — but their collective behavior amplifies or dampens upstream demand signals in ways that are not obvious from any single retailer’s policy.

Manufacturers face input cost volatility from the peanut commodity market and must decide when to pass costs downstream. Each manufacturer sets prices according to a rule that depends on margin, competitor pricing, and inventory levels. No manufacturer has full visibility into what competitors are doing — they observe prices with a lag.

Where It Gets Interesting

The non-trivial behavior emerges at the interfaces. Consider a commodity price spike — peanut prices rise sharply due to drought. Manufacturers face a margin squeeze. Some pass costs through immediately; others absorb the shock and wait.

In a standard equilibrium model, you would compute a new steady state. In the ABM, you observe the transition path: consumers who encounter price increases at one brand switch to store-brand alternatives, which drives down private-label inventory faster than anticipated, which triggers early reorders, which signals false demand to manufacturers. The bullwhip effect is visible directly in the simulation output without being imposed by assumption.

The 2008–2009 peanut paste contamination recall is the extreme version of this. A single supplier’s product contaminated a large fraction of downstream manufactured goods. The cascade — product recalls, retailer de-listing, consumer switching, manufacturer reformulation — played out over months and involved feedback loops across every layer of the supply chain. An ABM calibrated to the pre-recall market structure would have been able to stress-test exactly this kind of event before it happened.

What ABM Adds That Econometrics Does Not

Standard econometric approaches estimate average elasticities and use them to predict aggregate outcomes. They work well near equilibrium. They do not handle transition dynamics, inventory feedback loops, or heterogeneous responses across consumer segments that make supply chain shocks interesting.

ABM does not replace econometric estimation — it uses those estimates as inputs, at the individual agent level, and then simulates the system forward. The result is a model that can answer questions like: what fraction of consumers switch permanently after a price increase versus switching back when prices normalize? How long does it take for market share to stabilize after a recall? Which retailer reorder policy minimizes shelf-out events under demand uncertainty?

These are not hypothetical questions for peanut butter manufacturers. They are operational decisions made every quarter.

Takeaway

Agent-based models are not reserved for financial markets or epidemiology. Any market with heterogeneous agents, inventory dynamics, and competitive pricing produces the kind of emergent behavior that ABM is built to capture. Peanut butter is a clean, well-structured example — and the modeling lessons transfer directly to pharmaceuticals, electronics, and agricultural commodities where the stakes are considerably higher.

Building a supply chain simulation or market model? I’d be glad to help.