How rigid supply chains could turn balanced shocks into economic losses

From transaction maps to substitution possibilities

An input-output table records transactions among industries at a particular time. Satoshi Nakano and Kazuhiko Nishimura want to model something beneath that snapshot: the possibilities for changing input combinations when prices or productivity shift.

Their preprint, “Computing Endogenous Transformations in Processing Networks: A Dynamic Calibration Approach,” uses Cascaded Constant Elasticity of Substitution, or CCES, functions. These arrange inputs in a hierarchy, combining each input with an upstream composite. Substitution elasticities describe how readily those combinations can change in response to relative prices.

The distinction matters because a network with limited substitution cannot adjust as readily to a more expensive input. In the authors’ framework, changing prices can therefore reshape transaction shares rather than merely pass through an unchanged network.

Fitting a constrained network

The researchers anchor the model’s share parameters to a baseline input-output table and order inputs using an upstreamness measure derived from that table. They then calibrate substitution elasticities against historical transaction totals, treating observed prices as given during calibration.

Their objective balances relative errors in transaction predictions, a macro-aggregate discrepancy term, and a penalty pulling elasticities toward one. That last component favors the Cobb–Douglas benchmark unless the data support departures from it. Elasticities are constrained to be nonnegative, which the authors use to maintain the model’s economic concavity conditions.

Calibrating thousands of interacting parameters at once is computationally difficult. The proposed algorithm instead alternates across industries and, within each industry, adjusts one elasticity at a time, proceeding from upstream layers downward. It still uses the standard L-BFGS-B optimizer, but for individual scalar problems rather than the whole parameter matrix simultaneously.

In a 71-sector benchmark involving approximately 4,400 parameters, the authors report that the monolithic solver took more than 1,800 seconds to complete two initial iterations, reaching an objective value of 6.38. Their hybrid procedure reached 6.16 in roughly 1,000 seconds. These are reported computational comparisons, not evidence that every calibration problem will achieve the same speed advantage.

Checking productivity measurement

For a 10-sector U.S. application, the paper assumes structural parameters remain unchanged over a four-year window around the base year. Both optimization approaches reportedly reached a mean absolute percentage error of 1.65%, with macro-aggregate error reported as 0.00%.

The authors also compare productivity calculated from the calibrated CCES model with the conventional Törnqvist total factor productivity index. That index uses changing input cost shares and prices to measure productivity retrospectively. The paper reports close agreement across sectors during the calibration window.

This comparison supports the model’s ability to reproduce a familiar productivity measure within that window. It is not an independent demonstration of forecasting accuracy outside the calibration period.

When symmetric shocks produce asymmetric outcomes

The researchers next simulate 50,000 draws of normally distributed sectoral log-productivity shocks, with mean zero and standard deviation 0.20. They solve the network’s equilibrium price system and examine the resulting aggregate log-productivity outcomes.

Under the Cobb–Douglas benchmark, where substitution elasticities equal one, the simulated aggregate distribution is approximately symmetric: reported skewness is −0.01079. In the calibrated CCES network, skewness falls to −0.35560, indicating a more pronounced downside tail. The authors describe this distribution as having asymmetric, heavy-tailed risk.

Average outcomes also differ. Mean aggregate log productivity is −0.00068 in the benchmark and −0.00885 in the calibrated network, a difference of −0.00817 log units. A paired comparison across simulation draws finds the difference statistically significant at p < 0.01.

The authors call this “mean sinking”: within their model, symmetric microeconomic volatility lowers the expected aggregate log-productivity outcome relative to the benchmark. They interpret it as evidence of structural inelasticity. The result is a model-based comparison, not an observed estimate of an actual historical GDP loss.

What the findings leave open

The study is a preprint, and its conclusions depend on the calibrated structure and simulation assumptions. The authors acknowledge that value added is not separated into capital and labor, and describe the broad-sector application as a proof of concept. They also discuss noisy data and the impossibility of achieving an exact fit under their constraints.

The paper considers a further theoretical vulnerability: sufficiently severe negative shocks could eliminate a positive equilibrium. It explicitly does not claim that such a breakdown is imminent, and leaves its threshold unquantified. Its central finding is narrower: the calibrated network produces asymmetric downside risk and lower average log productivity under the specified shocks.