The Investment Acceleration Principle Revisited by means of a Neural Network

• Investment is among the most volatile and least predictable economic time series, despite being described as the engine of economic growth.Source: §1 · pp. 1–2
• The accelerator principle performs relatively well in empirical studies, but accelerator equations still fall short of reliable predictions.The principle models investment as dependent on variations in demand.Source: §1 · p. 2
• Standard business-cycle models assume firms invest proportionally to demand changes and observe demand only for their own products or sector.These assumptions imply decision-making myopia and help generate macroeconomic oscillations.Source: §1 · p. 2
• The principle was formulated to explain business-cycle expansions and crises, but recoveries are less clearly explained by demand-based autoregressive equations.The paper links the onset of recoveries to investments in novel fields opened by technological innovation.Source: §1 · p. 3
• The paper models firms’ decisions with a self-organising map whose neuron weights are coefficients of a disaggregated accelerator equation.The weights adapt the demand response to evolving technological possibilities.Source: §1 · p. 3

2 The Neural Network

• Each neuron models a firm in one of two sector layers: final-goods firms respond to demand and technology, while capital-goods firms respond only to capital-goods demand.A double feedback arises through the labour market.Source: §2 · pp. 5–6, Figure 3 · p. 7
• A Hebbian rule updates the coefficient matrices by strengthening weights associated with input-output combinations and applying forgetting over time.The stated rule uses μ for learning and ν for forgetting.Source: §2 · p. 9
• Combining the accelerator and coefficient-update equations maps technological information into disaggregated investment decisions, which can then be summed into aggregate investment.The accelerator coefficients are matrices A, B, and D.Source: §2 · pp. 8–9
• When a technological pattern emerges around mid-run, firms shift from erratic investment to an exponential growth path.The simulation feeds in a sinusoidal technology pattern that grows out of noise.Source: §2 · pp. 9, 11, Figure 4 · p. 10
• In all ten simulations, aggregate investments enter a stable exponential path between time steps 50 and 60.Figure 6 shows the ten runs and their average.Source: §2 · p. 11, Figure 6 · p. 12
• Recognition of innovations ends turbulence in the A, B, and D matrices, whose variation then follows an exponential path reflecting investment growth.Figure 7 plots the sum of absolute variations in each matrix, averaged over ten simulations.Source: §2 · p. 11, Figure 7 · p. 13

3 Analytical Description

• The analysis separates fast, basin-stable ∆c and ∆k from slow, unstable A and D; early recovery approximates fast adjustment, while plant-replication growth approximates fixed coefficients.The paper identifies these as two regimes of self-organising systems.Source: §3 · pp. 14–15
• Given technological information ∆e, the alternative algorithm derives ∆k, checks economic feasibility, then updates matrices A and D.Equation 8 yields the investment-change vector before feasibility rules are applied.Source: §3 · p. 15
• With information vectors assumed infinitely fast, aggregate investments settle at a constant value after a brief transition; this variant supplies A and D for a fixed-coefficient accelerator.The infinite-speed assumption removes production-time lags and information-flow effects that may cause excess productive capacity.Source: §3 · pp. 15–16
• After the technological pattern appears at step 30, A and D approach a stable path that can seed a fixed-coefficient accelerator with steeper exponential investment growth.In regime 2, these matrices are frozen at their step-30 values.Source: §3 · pp. 16–17
• Traditional fixed-coefficient accelerators account only for investments under an emerged technological paradigm; before that, the neural network improves performance.Figure 10 compares aggregate investments from the original and two-regimes models, averaged across ten simulations.Source: Figure 10 · p. 17, §3 · pp. 17–18

4 Concluding Remarks

• Kohonen neural networks offer a new way to represent economic agents’ decision-making and establish a theoretical bridge between an economic model and artificial intelligence.The paper presents this as an approach entirely new to economics.Source: §4 · p. 18
• The approach allows at least partial modelling of the otherwise vague concept of “corporate culture.”The paper also notes that the approach has practical implications.Source: §4 · p. 18
• Kohonen neural networks may help explain empirical investment dynamics, with an earlier attempt reported as encouraging.Source: §4 · p. 18
• Kohonen neural nets might use POS data to guide investment into novel fields, an unexplored application in the paper’s account.The paper says POS data had previously been used to forecast seasonal demand or demand at a new location.Source: §4 · p. 18