Hot Hand Diffusion Lab

Grounded in “Current fluctuations in a gas of active Ornstein–Uhlenbeck particles” — the paper shows that when a particle's velocity is itself a correlated (OU) process with persistence time τ, the variance of its time-integrated current is not stuck at the universal diffusive growth ∼t that ordinary (passive, memoryless) random walks obey. Depending on τ and the observation window, the scaled cumulant generating function that controls these fluctuations predicts super-diffusive growth (∼t², for windows short relative to τ) crossing over to ordinary diffusive growth (∼t) once the window is long enough to average over many correlation times.

This demo implements the real OU dynamics (exact discrete Ornstein–Uhlenbeck update, not a scripted "streak injector") for a swarm of bankroll walkers whose per-step edge is the particle's velocity. At τ=0 the fitted log-log slope sits at the standard diffusive exponent α≈1 — exactly what the "is the hot hand real" skeptics assume when they treat streaks as noise and size bets with Kelly's sqrt(time) variance rule. Raise τ and the fitted slope visibly bends toward 2: your bankroll's fluctuations are no longer diffusive, Kelly's implicit variance assumption breaks, and a real hot streak (serial correlation in the edge itself) can make a "lucky" run statistically expected rather than an outlier — the same math that makes the active-particle current non-universal.