polymarket-kernel

Ultra-low latency computational core for Polymarket market making, now upgraded into a comprehensive decision-support and risk engine for prediction-market microstructure.

Repository Layout

• packages/crates: Rust crate (polymarket-kernel)
• packages/npm: public npm package
• packages/bun: public Bun package
• packages/python: public PyPI package
• docs: additional project documentation
• tools: local release and utility scripts

Published Packages

• crates.io: polymarket-kernel
• npm: holypolyfoundation-bs-p-npm
• Bun (npm registry): holypolyfoundation-bs-p-bun
• PyPI: bs-poly

Install

# Rust
cargo add polymarket-kernel
Node.js / npm
npm i holypolyfoundation-bs-p-npm
Bun
bun add holypolyfoundation-bs-p-bun
for native build in Bun, trust install scripts for this package:
add "trustedDependencies": ["holypolyfoundation-bs-p-bun"] to package.json
then run bun install
Python
pip install bs-poly

Usage Guide

Rust

use polymarketkernel::calculatequotes_logit;
fn main() {     let x_t = vec![0.1];     let q_t = vec![0.0];     let sigma_b = vec![0.2];     let gamma = vec![0.08];     let tau = vec![0.5];     let k = vec![1.2];
let mut bid_p = vec![0.0; 1];     let mut ask_p = vec![0.0; 1];
calculatequoteslogit(&xt, &qt, &sigmab, &gamma, &tau, &k, &mut bidp, &mut ask_p);     println!("bid={:.6}, ask={:.6}", bidp[0], askp[0]); }

Node.js / npm

import { sigmoid, logit, calculateQuotesLogit } from "holypolyfoundation-bs-p-npm";
console.log(sigmoid(0)); // 0.5 console.log(logit(0.5)); // 0
const out = calculateQuotesLogit([0.1], [0.0], [0.2], [0.08], [0.5], [1.2]); console.log(out.bidp, out.askp);

Bun

import { sigmoid, logit, calculateQuotesLogit } from "holypolyfoundation-bs-p-bun";
console.log(sigmoid(0)); // 0.5 console.log(logit(0.5)); // 0
const out = calculateQuotesLogit([0.1], [0.0], [0.2], [0.08], [0.5], [1.2]); console.log(out.bidp, out.askp);

Python

import bs_p
print(bs_p.healthcheck()) print(bs_p.sigmoid(0.0))  # 0.5 print(bs_p.logit(0.5))    # 0.0
out = bsp.calculatequotes_logit([0.1], [0.0], [0.2], [0.08], [0.5], [1.2]) print(out["bidp"], out["askp"])

Overview

polymarket-kernel implements a unified logit-space stochastic framework where probabilities are transformed into log-odds and processed through SIMD-native math.

The crate now combines:

• high-throughput quoting primitives
• inventory-aware execution math
• vectorized analytics and portfolio risk aggregation

Source paper:

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

Features

Core Quoting Kernel

• SoA (Structure of Arrays) layout for contiguous memory access and SIMD-friendly loads
• Runtime-dispatched AVX-512 quote acceleration with portable fallback
• Native NEON (aarch64) path with fully vectorized exp/log1p/sigmoid/logit (~1-2 ulp)
• Inventory-aware Avellaneda-Stoikov quoting in logit space
• Exact, numerically stable sigmoid/logit mapping across the public API (logit keeps full relative accuracy near p = 0.5 via a log1p formulation)

Analytics Capabilities

• Implied Belief Volatility calibration from market bid/ask quotes (exact closed-form inversion, no iteration)
• Vectorized Stress-Testing (what-if analysis) for shocked probabilities, PnL shifts, and re-quoted books
• Adaptive Kelly sizing for maker and taker clip recommendations under inventory and risk constraints
• Order Book Microstructure metrics: OBI, VWM, and pressure signal in logit space
• Cross-Market Portfolio Greeks aggregation with optional weighting and correlation matrix support

Systems Properties

• C kernel in packages/crates/c_src/ with FFI-safe Rust bindings in packages/crates/src/
• Portable scalar baseline that runs on any CPU
• Zero allocations in the hot path (pre-allocated caller-managed buffers)
• Runtime-dispatched AVX-512 fast path on supported server-class CPUs; compile-time NEON fast path on aarch64
• Shared SIMD math header (csrc/pmsimd_math.h) with vectorized exp/log/log1p/sigmoid/logit used by both quoting and analytics
• Numerically safe clamping for stable logit evaluation without saturating large logits
• Lock-free SPSC ring buffer for market data handoff with cached-index design (producer and consumer avoid touching each other's cache line on the hot path)

Quick Start

Install:

cargo add polymarket-kernel

Call calculatequoteslogit with SoA input slices:

use polymarketkernel::calculatequotes_logit;
fn main() {     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];
let mut bidp = vec![0.0; xt.len()];     let mut askp = vec![0.0; xt.len()];
calculatequoteslogit(         &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}", bidp[i], askp[i]);     } }

Analytics API Example

use polymarket_kernel::{analytics, GreekOut};
fn main() {     let n = 4usize;
let bid_p = vec![0.49, 0.41, 0.62, 0.23];     let ask_p = vec![0.52, 0.45, 0.66, 0.27];     let q_t = vec![8.0, -4.0, 2.0, 0.0];     let gamma = vec![0.08; n];     let tau = vec![0.5; n];     let k = vec![1.4; n];
let mut implied_sigma = vec![0.0; n];     analytics::impliedbeliefvolatility_batch(         &bid_p,         &ask_p,         &q_t,         &gamma,         &tau,         &k,         &mut implied_sigma,     );
// q_t is retained for API-shape consistency, but the current calibration     // formula depends on spread, gamma, tau, and k.     let x_t = vec![0.20, -0.40, 0.70, -1.10];     let shock_p = vec![0.01, -0.02, 0.03, -0.01];
let mut outrx = vec![0.0; n];     let mut out_bid = vec![0.0; n];     let mut out_ask = vec![0.0; n];     let mut out_greeks = vec![GreekOut::default(); n];     let mut out_pnl = vec![0.0; n];
analytics::simulateshocklogit_batch(         &x_t,         &q_t,         &implied_sigma,         &gamma,         &tau,         &k,         &shock_p,         &mut outrx,         &mut out_bid,         &mut out_ask,         &mut out_greeks,         &mut out_pnl,     );
println!("implied sigma: {implied_sigma:?}");     println!("stress pnl shift: {out_pnl:?}"); }

Benchmark Snapshot

Apple M4 (aarch64, NEON path):

============================================================
 POLYMARKET-KERNEL RAW BENCHMARK
============================================================
 Quote Batch Size        :       8192 markets
 Quote Iterations        :     100000
 Runtime-Dispatch Quote :       5.78 ns/market
 Sigmoid Batch           :       1.20 ns/element
 Implied Vol Calibration :       5.78 ns/market

 SPSC Ring Capacity      :    1048576
 SPSC Messages           :   10000000
 SPSC Throughput         :     188.42 M msgs/sec
============================================================

For reference, the same host on the previous scalar-fallback build measured 11.07 ns/market for quotes and 55.9 M msgs/sec for the SPSC ring (~1.9x and ~3.4x respectively).

License

MIT