Lucas Pirvuti

Limit Order Book Sim

Reconstructed a 10-level order book by replaying 455K Nasdaq ITCH messages, then validated it against the exchange's own book over 170K comparisons with zero mismatches on fully observed levels. Standard library only.

• Fill simulator modelling real queue position across 5 latencies and 3 cancellation models
• Measured passive quoting losing 0.8 cents per share against a 0.2 cent rebate
• Latency cut the fill rate from 83% to 6% — the cost of being slow, quantified

Green Simulation Limits

Reproduced a 2022 Monte Carlo method that reuses samples across scenarios and confirmed its 100× error reduction — then found the advantage vanishes by eight dimensions once the density it depends on has to be estimated from data.

• Error measured against a closed-form benchmark, itself checked on 2M brute-force Monte Carlo draws
• Control estimators blind to the conditioning features confirmed dimensionality caused the loss
• Negative result written up as the finding rather than tuned away