shubhxho/polymarket-hft-benchmarks

Rust paper-engine benchmarks

The local research desk exposes simulation playback, supplied-candidate ranking, benchmark provenance and retained local inference. Terminal GIF. Default operation has no paid API calls. Browser interaction/layout checks are unverified; no browser was connected and local TCP binding was refused by the validation sandbox. CPU API dispatch and packed-model inference were tested through the command-line transport. No company affiliation is claimed.

Reproducible local measurements for the Rust paper engine and optional Polygon research planner. These are synthetic paper-engine and kernel benchmarks, with no live orders, historical alpha evaluation or fund-return claim.

The dataset contains 22,500 raw timing batches from five rounds of equal-output top-of-book lookup. Three implementations—our cached lookup, a full-array scan and a BTree reference—query the same 64 synthetic books at 1, 16 and 128 levels per side. Each timed batch contains 512 lookups; each backend gets 100 warmup batches and 500 measured batches per round. Backend order rotates.

Measured wins and mixed results

At 16 and 128 levels per side, cached lookup was 1.33–1.40x faster than BTree top lookup across all five rounds. Compared with a full-array scan it was approximately 202–214x faster at those depths. These ratios concern lookup only: maintaining the cache, updates, matching, network I/O and persistence are excluded.

At one level per side, BTree comparison was mixed (0.35–1.20x); there was no consistent win. See the full comparison report, raw comparison JSON and environment. Ranges describe observed round-to-round variation, not confidence intervals.

Packed affine q2 weights used 384 bytes versus 4,096 bytes for the synthetic dense f32 vector: 90.63% less storage, including packing/scales/biases. Packed reference kernels were slower than dense f32 here; memory savings are not a speed or trading-quality claim.

Current Polygon revision

Version 0.4.0 migration evidence is in POLYGON_MIGRATION.md. Its 29-row compact CPU benchmark is labeled separately and cannot be used as a speed comparison against the historical release results below. CUDA remains uncompiled and unmeasured on this host.

Contents

• data/lookup-samples.csv: viewer-ready raw batch data.
• reports/comparison.json: all rounds and raw samples with output-equality checks.
• reports/benchmark-after.json: default optimized Rust operation measurements.
• reports/benchmark-mlx-cpu.json: native MLX CPU release kernels, input creation/evaluation/readback included.
• reports/BENCHMARKS.md: timings and storage/error tables.
• reports/polygon-check.txt and reports/run-through.txt: current Polygon correctness and CLI evidence. Earlier release logs describe source version 0.3.0.
• docs/: architecture, operations, schemas and research methods.
• source.tar.gz.base64 and REPRODUCE.md: reproducible Rust source bundle and its SHA-256.

Columns

Column	Meaning
levels_per_side	Number of occupied levels on each side
round	One of five independently timed rounds
backend	cached, full_scan or btree
sample_index	Batch index within the round/backend
batch_size	512 lookups
batch_ns	Wall-clock nanoseconds for the entire batch
amortized_ns_per_lookup	batch_ns divided by batch_size; not individual-request latency

Validation and limits

Current Polygon version 0.4.0 passed 46 compact-profile tests; one NVIDIA-only parity test is explicitly ignored. Historical version 0.3.0 passed 42 default and 44 native MLX CPU tests. Formatting, Clippy, replay, reversal, journal restart, news rejection, signal math, paired scoring, code generation and installation were exercised.

The Rust signal interface exports expiring candidates with exact fee/edge math, conservative probabilities and capped sizing. The weather baseline now uses conditional truncated-normal normalization, verified against a half-normal reference. The paired scoring fixture is synthetic and does not establish FPT superiority. See docs/SIGNALS_AND_MATH.md and reports/FPT_STATUS.md.

The host is macOS ARM64 with Rust 1.97.1 / LLVM 22.1.6, fat LTO and one codegen unit. CPU model, RAM and core count could not be captured because sandbox sysctl access was denied. No CPU pinning or thermal isolation was used.

Metal GPU is unmeasured because its compiler is unavailable. Jev timings use synthetic recorded responses; live API accuracy, latency and cost are unmeasured. No whole-system comparison was run against HRT, hftbacktest, Barter or the official venue client.

The current Rust engine, signals and planner enforce Polygon 137; Base inputs and historical Base journals reject. Historical version 0.3.0 release measurements remain Base-tagged synthetic research results, not new Polygon or CUDA measurements. Signing and live execution remain unavailable. Declared news source URLs are not authenticated. This dataset supports reproducible research; it does not establish live fund readiness.

Project: GitHub. Public mirrors: benchmark Gist.

The Rust pipeline now includes bounded reusable I/O, fixed 40-byte local updates, corruption-checked HQT1 weights, explicit fractional Kelly and an opt-in Polygon unsigned planner. CUDA source is experimental and unmeasured. Neither Polymarket nor Robinhood authenticated trading is implemented. See docs/TRANSPORT_AND_BACKENDS.md and docs/POLYGON_EXECUTION.md.

The simulation extension adds 336 deterministic synthetic paths and 672,000 engine invariant checks. Losses are retained, including a negative normal/aggressive median; no profitability claim follows. See reports/SIMULATIONS.md and docs/SIMULATIONS.md. Current CPU test gate: 46 passed, one NVIDIA-only test ignored. CUDA waiting/stream changes remain GPU-unverified.

$10,000 synthetic weather portfolios

Six scenarios, 96 independent seeded portfolios, 30,720 contracts and 153,600 bucket evaluations. Normal median P&L was -$117.71; forecast bias median was -$1,041.56. No profitability is established. Full methods and all outcomes, raw report, animated release evidence. Generator source is bundled; synthetic forecasts/resolutions can be regenerated deterministically. Native MLX checks for this update were blocked by disk preflight. CUDA and live execution remain unvalidated.

Rust streaming update

Bounded streams and accelerator evidence: 56 CPU tests, 59 native MLX CPU tests and Clippy pass. Equal-output batching reduced in-memory writer calls from 128 to one for the 128-record fixture. This is not a network/HFT speed claim. Raw stream benchmark, compact native CPU benchmark. Metal and CUDA remain unvalidated.