Elia Zonta

Graduate researcher under the supervision of Prof. Tomaso Poggio, focusing on the mathematics of deep learning and theory of intelligence. Having fun with self-stabilization models, multi-indexes and edge of stability.

Focused on the intersection of network science, financial theory and systemic risk.

Coursework: Stochastic Processes, Computational Statistics, Deterministic & Stochastic Optimization.

Tuition merit & Research Thesis scholarship. Attended MS/PhD level computer science and math courses as a 1st & 2nd year undergraduate.

Mechinterp ? Agents ? Knowledge Graphs ? MCPs ?

Built a fully automated sports arbitrage system using Python, Selenium, and statistical modeling to scrape real-time odds from 5+ major sportsbooks, detect market inefficiencies, and execute 200+ monthly trades with a 6.2% theoretical ROI.

Eleven classifiers across five tiers on the UCI Spambase corpus, built to answer not only which model wins but which were worth training at all. LightGBM tops the leaderboard (test F1 0.9407, 10-fold CV 0.9483 ± 0.0074), but the interesting result is the audit: the four tree models' errors correlate at 0.71–0.85, which predicts before running it that the soft-vote ensemble scores below its own best member. An accuracy-vs-fit-cost analysis separates genuinely better models from merely more expensive ones, trimming eleven candidates to a five-model roster. Written as a reactive marimo notebook, and explicit about the 391 duplicate rows that inflate every published Spambase score — including its own.

A depth- and fee-aware scanner for cross-outcome ("dutch-book") arbitrage in Polymarket's multi-outcome markets, showing that the apparent edge does not survive taker fees and real order-book depth and quantifying where the durable edge actually sits on the maker side: breakeven fill time from measured volatility, the liquidity-reward subsidy, and risk-free basket inventory netting. Built on live CLOB order books with a tested analytics core.

Measures how fast an order-book signal decays on real NASDAQ limit-order-book (LOBSTER) data. Microstructure feature engineering runs natively in kdb+/q, streamed to Python over a PyKX IPC bridge, with a LightGBM model per prediction horizon. Emphasis on leakage-aware evaluation — purged walk-forward splits, moving-block bootstrap confidence intervals, and guards against degenerate targets and mechanical micro-price artifacts.

A low-latency trading engine in C++20: lock-free single-producer/single-consumer ring buffers, cache-line-aligned data structures, integer-only (zero-FPU) technical indicators, and zero-copy telemetry streamed to a Python dashboard over a memory-mapped ring buffer. Runs a full tick -> bar -> indicator -> strategy -> execution pipeline across dedicated threads on simulated or replayed market data.

Simulation of the auto-deleveraging cascade triggered by the May 2021 flash crash on the Bitcoin market. Custom simulator for margin trading and auto-deleveraging mechanisms, focusing on replicating the dynamics of the event and analyzing systemic risk implications. Implemented in Python with detailed documentation and visualizations.