Hikaru Saijo

Today I'm presenting my new paper, “Macroeconomic Effects of Production Networks: A Deep Learning Approach,” at the SI 2026 Impulse and Propagation Mechanisms meeting, with Christian Matthes (@cmatthes_econ), Hikaru Saijo (@hpfilter), and Felipe Schwartzman. The talk focuses on the substantive (and more interesting) findings about the dynamic properties of production and investment networks, but it builds on the Deep Sets architecture I developed with Mahdi Kahou (@MahdiKahou). That work leans on the symmetry group (a great unknown in economics, unlike in physics) and on the powerful implications of symmetry and permutation invariance. There will also be a brief cameo of non-removable singularities in perturbation solutions, a point that, unfortunately, many macroeconomists do not fully appreciate. If you are at the NBER SI, come by. It might be fun. And @FlorinBilbiie will present a really cool paper after me.

Saturday, I gave a talk at ASSA 2026 on my new paper ``Macroeconomic Effects of Production Networks: A Deep Learning Approach,’’ with @cmatthes_econ, @hpfilter, and @MacroInPieces. Several attendees asked for a copy of the slides I used (sorry, we are still a few weeks away from circulating a completed draft), so I am linking them here: sas.upenn.edu/~jesusfv/Slide… There has been a lively debate over the last few weeks on X about what deep learning can do in macro. I think this is a good example of what we can now study. Dynamic stochastic models with non-trivial input-output structures are a beast to compute. They are inherently nonlinear, and you need to carry not only aggregate state variables but also all the sector-specific variables, such as sectoral productivity and sector-specific capital. If you have, as we do in the paper’s calibration to match U.S. input-output tables, 37 sectors, you end up with at least 74 state variables. Even very aggressive sparse-grid allocations, such as Smolyak, will struggle with that. A deep neural network can deal with this problem, especially if you build it using the symmetry (or, if you prefer, exchangeability) ideas I discuss here: nber.org/papers/w28981 with @MahdiKahou, @jlperla, and Arnav Sood. We also show that adopting a deep learning approach changes the answers you get relative to more traditional solution methods. More concretely, we capture endogenous changes in relative prices across sectors that perturbation solutions largely miss. That leads to materially different elasticities of output, investment, and labor supply to sectoral shocks. In our benchmark calibration, we find substantial attenuation of sectoral shocks, precisely because of these relative price changes. When steel becomes too expensive following a productivity shock in the steel industry, car manufacturers switch to more plastic. This result resonates with my own earlier work on the macroeconomics of wars. One lesson from wars over the last 120 years is how resilient modern economies have been to large disruptions from mobilization, blockades, or bombing. What looked like bottlenecks that would stop production turned out to be inconveniences, because there are many ways to produce a tank or an airplane. Some alternatives were less efficient than the optimal one, but the productivity losses were not as large as pre-war planners envisioned when they took a static view of resource allocation. Our results replicate that intuition inside a standard stochastic multi-sector business cycle model. Of course, we could throw wrenches into the economy’s reallocation process, but, at this moment, I think the insight is robust.