Quantitative research · AI evaluation · Vancouver
Tave La Rocque
UBC engineering student on full-time research leave, working on market microstructure, online calibration, and whether AI benchmarks measure what they claim to.
Research
05Order-driven volatility fails one of four predictions out of sample
A pre-registered test of Maitrier and Bouchaud's 2026 framework on 8 full NASDAQ days, with code frozen before any held-out stock was touched. One prediction is rejected in both tick-size bins and in every confirmatory cell; the other three hold. The run reproduces bit for bit across 2,823 fields. Paper in preparation.
Exact minimax values for an open problem in online calibration
Exact values of the sign-preservation-with-reuse game behind the sequential calibration bounds of Dagan et al. (STOC 2025), certified in both directions: lower bounds from exact rational strategies, upper bounds from Farkas certificates. A 33-page single-author paper, forthcoming.
An AI cheating benchmark feeds models a SyntaxError on 1 in 5 attempts
A pre-registered re-scoring of 714 ImpossibleBench episodes across two models. The benchmark's extraction rule shows the model a SyntaxError instead of test results on about 20% of attempts, and requiring tests to finish cuts one model's measured cheating rate from 59.2% to 19.4%.
Implied Odds: prediction markets against option-implied probabilities
Compares Kalshi contracts on the S&P 500, Nasdaq-100 and Bitcoin with Breeden-Litzenberger densities from arbitrage-checked SVI smiles. Collects data around the clock; every test is pre-registered and judged on a locked holdout.
Look-ahead leakage in point-in-time LLMs
Do language models trained only on data up to a date really know nothing after it? An in-progress, pre-registered audit of whether point-in-time LLMs leak post-cutoff knowledge into financial forecasts.
Builds
06itchbook
NASDAQ's full order-by-order book rebuilt from 423M raw TotalView-ITCH 5.0 messages a day, output byte-identical to an independent Python reference. Found that 85% of cancelled volume leaves from behind a resting order, so standard L2 queue models overfill backtests.
protocol-emulator
A two-thread programmable protocol CPU for Jane Street's ASIC competition, designed for Tiny Tapeout's IHP 130 nm shuttle and passing signoff. Formally verified pins to pins against an independent spec with a SAT equivalence proof.
holdem-cfr
A counterfactual regret minimization solver for heads-up no-limit hold'em, verified against exact best responses at 4.9×10−6 exploitability on Leduc. Multithreaded results are bit-identical to serial.
ocaml-contracts
Peyton Jones, Eber and Seward's Composing Contracts as a typed DSL, pricing Kalshi Bitcoin contracts from Deribit-implied densities, with Greeks, delta hedging and an Avellaneda-Stoikov market maker.
darkmpc
A dark pool even its operator can't see into. Orders are secret-shared across three servers that run the auction jointly, so no single server ever sees an order.
MODUX
A training app for hybrid athletes, co-founded and solo-built end to end. Its readiness score models the conflict between strength and endurance load as an 11-dimensional fatigue vector with 72-hour decay.
Experience
04Vancity · Indigenous Project Assistant Intern
Selected 47 of 8,500 applicants for the inaugural cohort. Rebuilt a complaints risk classification system across 2,370 cases; its validation layer caught a 10× reporting error before executive review.
Current at UBC · Co-founder & CEO
One evening in April 2027 putting 300 UBC students in a room with Vancouver's finance and venture leaders. Three keynotes who stay for the networking floor; all net proceeds to BC Children's Hospital Foundation.
KS Society · Director of Philanthropy
Promoted after leading external recruitment; owns philanthropic strategy and partner outreach.