“First, we extend the state space to include equilibrium prices and model parameters, which allows us to clear markets and estimate parameters by solving the model once”
This is exactly what the toolkit does with “joint-optimiazation”, right?
Paper is an exercise in obfuscation.
As you say they claim four contributions, one of which is “First, we extend the state space to include equilibrium prices and model parameters, which allows us to clear markets and estimate parameters by solving the model once”, which is just joint-optimization and is completely standard (has been around since last millennia).
The rest of the exercise is surrogate optimization. They build the surrogate as a neural-network. Then they estimate the model using the surrogate.
Likely something is wrong with their codes though. They claim they can train the surrogate model of the entire parameter space in 20 minutes —which would typically require thousands of model solutions, unless the space is degenerate— but to estimate the model —which would typically require thousands of model solutions— they require days. How you can do one in way less time than the other when both require you to solve the model thousands of times strongly points to one of the two being a mistake.
By comparison, if you look at Chen, Didisheim & Scheidegger (2026), who also build surrogate to do estimation, it makes sense in their instance as they build the surrogate once and then use it to estimate the model a few thousand times (a different estimate for each day of financial data). So even though building the surrogate takes way more model solutions than a single model estimation would require, it still works out to be more than worthwhile because the one surrogate is then used for many estimations.
There is no AI anywhere in the whole paper, except right near the end where they say “We have verified that the agent can apply our method to simple problems ranging from a deterministic cake-eating problem to a stochastic growth model with four estimated parameters. We have not yet tested more complex models.” They certainly don’t appear to let AI anywhere near the actual exercise that the paper is about.
PS. The surrogate estimation using a neutral-net exercise itself is cool. But the write up is atrociously misleading and it does seem implausible that their runtimes are quite so different when both approaches should require solving the model a lot of times. Certainly the paper needs an explanation of why surrogate estimation is so much faster in this instance, my understanding is surrogates can only be faster when solving the model itself is very time consuming and so you want to reduce doing this.
1 Like