Quantitative research · PhD, UC Santa Barbara
Quantitative researcher at Wells Fargo in Charlotte. I work on game theory, empirical Bayes, machine learning, and quantitative finance, and increasingly on high-dimensional probability and statistics. My PhD work at UCSB solved the problem of forecasting the forecasts of others, which economics has had to work around since Keynes. Here I share projects and writing about what’s been on my mind. Feel free to reach out!
New here? Start with the noise-state, explained, or solve a game in your browser.
the noise-state: what player i believes, at time t, about the shock at u
Actions move the state, the state sets what others observe, and what they observe sets their next actions. My dissertation keeps that loop intact by tracking beliefs about the primitive shocks. Move a slider and the equilibrium is solved again in about a second.
Solve a game →
Thousands of coins, each flipped a few times, and odds that change somewhere on the square. The estimator adds detail only where a false-discovery gate says the flips justify it, then stops on its own. It separates the odds from each coin's own effect and from the noise of the flips. Draw the odds yourself and see what it can find.
Draw your own odds →Freeform triangles, right triangles, rectangles and a regression tree, grown on the same noisy surface one greedy cut at a time, with the same number of free parameters each.
Watch four meshes grow →Why knowing what others know changes everything, starting from a doctor’s diagnosis and the sun exploding, and how the dissertation handles beliefs about beliefs.
Start with the story →The information wedge: what it is worth to move an opponent’s beliefs, why it vanishes when the loop is cut, and why starving one player of information can help everyone.
See the wedge solved →How the decision mesh decides where the data justify more detail, and when to stop, drawn from a live run in your browser.
Follow one fit →One level further down: how a split spreads, how a candidate is scored and corrected, how the null is fitted, and how the two engines differ, drawn from live runs in your browser.
Go through the algorithm →What the explorer does when it solves a game: kernels on a grid, one best response, the fixed point and its acceleration, the checks, drawn from solves run in your browser.
Watch a solve →My dissertation. Each player keeps estimates of the primitive shocks instead of belief hierarchies, so beliefs, prices and policies are impulse responses and equilibrium is a fixed point in them. The information wedge prices changing an opponent's beliefs.
Exact equilibrium characterization of finite-player continuous-time LQG games with endogenous signals
Adaptive, continuous regression that makes grounded, systematic decisions about where to refine the surface, and decomposes the data into the surface, unit-level effects, and pure noise. The rectangular version splits one axis at a time, like a decision tree, so it scales to higher dimensions. It's also very interpretable, so it's easy to tinker with.
