Jay Forrester, 1950s — computing a system's behaviour from four parts: stock, flow, feedback, delay
Referenced by: Topic 6 Feedback Loops
01The Question It Poses
In the 1950s, Jay Forrester at MIT crossed from engineering control into the study of firms and hit a recurring oddity: a company's orders, inventory, and headcount would boom and bust periodically, while management could always find a "cause" for each swing (some customer, some promotion, some decision) — yet swap the people, hold the meeting, and the swings kept coming.
Forrester's question was: is there a way to skip chasing the "who and why" behind each swing and instead write down how a few internal loops are wired, then compute directly how the system must move? System dynamics is his answer — a language that translates "structure" into equations and runs out "behaviour." Its bet: most persistent dynamics (growth, collapse, oscillation) come not from outside but are endogenous to structure.
02The Rules
The whole language has just four basic parts; every complex model is built from them:
Stock: the amount that accumulates — the system's "memory." Water in a tank, money in an account, staff on the payroll. Drawn as a box.
Flow: the rate that changes a stock per unit time — the only thing that can change a stock directly. Inflow/outflow, income/spending. Drawn as a pipe with a valve.
Feedback: the stock's current value loops back to affect a flow. Level high, so release less (negative/balancing); more money, so more interest (positive/reinforcing). Drawn as an arrow with a sign.
Delay: feedback takes time to loop back. Information delay (you learn late) or material delay (things in transit). Drawn as a hash across the arrow.
The core mathematical fact is dead simple: a stock's rate of change at any instant = inflow − outflow. The stock itself is that difference accumulated over time (an integral, mathematically). That one rule, plus "flows are in turn shaped by stocks through feedback loops," closes the four parts into a ring, and the behaviour is locked in.
The standard notation of system dynamics. Any complex model — a company, a city, a forest — is these four parts assembled over and over.
03What You See When It Runs
Strikingly, there are only a handful of ways to wire loops, so there's only a handful of behaviour "curve families." Learn these few shapes and you can almost read structure back off behaviour at a glance:
The four basic behaviour curves. Growth points to positive feedback, stability to negative, an S-curve is "positive first, negative second" handing off, and oscillation always hides a delay.
Memorize these four and many real curves become "structure you can read": an S-curve tells you a positive-feedback engine hit some negative-feedback ceiling (market saturation, resource exhaustion); a sustained oscillation tells you the loop must hide an unnoticed delay. The shape of behaviour betrays the wiring of structure — which is system dynamics' most practical gift: inferring the invisible loop from the visible behaviour.
04What It Explains
System dynamics doesn't predict a specific number on a given day; it explains why a whole class of systems shows a certain typical dynamic shape, and lets you test in a model — before you act — which lever actually changes the behaviour. Its classic arenas:
the supply chain's bullwhip effect (a small twitch at the end → big swings upstream), Forrester's "Urban Dynamics" (why some aid policies actually speed decay), the Meadows team's globally-debated 1972 Limits to Growth (population, resources, and pollution wired into one global loop), and the growth-investment-capacity cycle inside firms. Its core insight is one sentence, but a hard one: counterintuitive behaviour is usually the inevitable product of a perfectly understandable structure — you need not assume anyone is stupid or malicious; wire the loop that way and the outcome is fated.
What It Cannot Explain
It can't fix your parameters for you. A loop diagram is qualitative — it tells you "it will oscillate," but how large, what period, where the tipping point sits, all depend on numbers you often can't measure well. A pretty causal diagram gives the illusion of "I get it," but without reliable parameters it's far from predictive.
It can't handle systems whose agents learn and game. Classic system dynamics treats loops as relatively fixed mechanical pipes. But when a system is made of people who anticipate the loop and act ahead of it (financial markets, arms races), the pipe itself shifts with expectations — game theory or agent-based modelling fits better there.
It can't explain systems driven mainly by external shocks. Its strength is endogenous dynamics. If a thing's ups and downs come mostly from unrelated random events crashing in from outside, drawing its internal loops is beside the point. Test: turn off the external inputs — does it still move on its own?
It can't substitute for spatial and network structure. Standard stocks and flows treat everything in a stock as one uniform pot. But an epidemic spreading over a human network, or resources distributed across geography, behave in ways that depend strongly on "who's connected to whom" — that needs a network model, not one well-mixed tank.
Where you draw the boundary is where the conclusion skews. A model must fence in "which loops count and which don't." That boundary is drawn by the modeller, and once an important external loop is left outside, the model will confidently give a systematically wrong conclusion — and it won't tell you what it left out.