polymarket-kernel

Ultra-low latency computational core for Polymarket market making, based on the Logit Jump-Diffusion framework and an Avellaneda-Stoikov adaptation in logit space.

Overview

polymarket-kernel implements a unified stochastic kernel for prediction markets where probabilities are transformed into log-odds and processed with vectorized math.

The crate is designed for:

• HFT backtesting engines
• live market-making bots
• inventory-aware quoting across large market batches

It exposes an FFI-safe Rust API backed by a C SIMD kernel.

Source paper:

• Toward Black-Scholes for Prediction Markets (Shaw & Dalen, 2025)

Features

• SoA (Structure of Arrays) layout for contiguous memory access and SIMD-friendly loads
• AVX-512 vectorized quote engine for batch processing
• Custom AVX-512 log1p approximation to avoid scalar fallback in spread computation
• Fast sigmoid approximation for x -> p mapping in hot paths
• Inventory-aware Avellaneda-Stoikov quoting in logit space
• Zero allocations in the hot path (pre-allocated input/output buffers)
• Numerically safe clamping for stable logit/sigmoid evaluation

Quick Start

Install:

cargo add polymarket-kernel

Call calculate_quotes_logit with SoA input slices:

use polymarket_kernel::calculate_quotes_logit;

fn main() {
    // SoA inputs for N markets.
    let x_t = vec![0.15, -0.35, 0.90, -1.20];
    let q_t = vec![10.0, -6.0, 3.0, 0.0];
    let sigma_b = vec![0.22, 0.18, 0.30, 0.15];
    let gamma = vec![0.08, 0.08, 0.08, 0.08];
    let tau = vec![0.50, 0.50, 0.50, 0.50];
    let k = vec![1.40, 1.25, 1.10, 1.80];

    // Pre-allocated outputs (no hot-path allocation inside the kernel).
    let mut bid_p = vec![0.0; x_t.len()];
    let mut ask_p = vec![0.0; x_t.len()];

    calculate_quotes_logit(
        &x_t,
        &q_t,
        &sigma_b,
        &gamma,
        &tau,
        &k,
        &mut bid_p,
        &mut ask_p,
    );

    for i in 0..x_t.len() {
        println!("market {i}: bid={:.6}, ask={:.6}", bid_p[i], ask_p[i]);
    }
}

License

MIT