An empirical research testbed, execution zoo, and parametric null benchmark for quantitative alpha models. Reviving historical model artifacts at an identical prospective starting line under standardized finite-capital execution constraints.
QuantPits Arena is an empirical evaluation framework, benchmark zoo, and interactive visualization testbed.
- Live Platform: Access the hosted arena directly at arena.quantpits.com.
- 100% Offline Ready: The web research platform (
web/) is completely self-contained and pre-packaged with complete tournament simulation data:- Zero external package dependencies (native Python standard library server).
- Complete coverage of 168 contestant-animal execution paths, 11,000 parametric monkey null simulations, cross-model trajectory curves, counterfactual decision audits, and behavioral fingerprints.
# Clone the repository
git clone https://github.com/QuantPits/QuantPits-Arena.git
cd QuantPits-Arena
# Launch local exploration server
python3 -m http.server 8080 --directory web
# Open in your browser
# http://localhost:8080- Visual Exploration vs. Raw Execution:
- Offline Analytics (Out-of-the-Box): The web testbed comes with pre-compiled tournament datasets, allowing immediate visual exploration and cross-model comparison without any training or inference setup.
- Running Custom Evaluations: If you wish to run the evaluation pipeline from scratch or benchmark your own custom models through the Zoo:
- User-Provided Models: Model weights are not bundled in this repository. Users must supply their own local model artifacts or checkpoints (conforming to the configuration manifests in
manifests/public/). - User-Provided Market Data: Historical market quote feeds and feature matrices for your target stock universe and evaluation dates must be provided locally.
- Included Components: This repository supplies the 28 animal execution policies, finite-capital transaction and portfolio accounting engine, parametric monkey null generators, and report export tooling.
- User-Provided Models: Model weights are not bundled in this repository. Users must supply their own local model artifacts or checkpoints (conforming to the configuration manifests in
- Not a Live Trading Engine: The repository is an empirical evaluation testbed; it does not provide real-time order routing, brokerage gateway connectivity, or continuous live execution infrastructure.
- Declarative Manifests: Strategy models are specified as reproducible configuration manifests (
manifests/public/) documenting architectural parameters, feature groupings, and standardized inference adapters.
- Official Evaluation Baseline: 2026-07-03 ~ 2026-08-28 (41 trading days, Weeks 1–8).
- Evaluation Nature: Retrospective backtest baseline simulated with knowledge of July–August market conditions to establish initial tournament standings.
- Official Release: Complete 168 execution paths, parameter distributions, and Episodes 01–08 Dispatch.
- Current Standing: High selection intensity dominated during the summer momentum regime, with
CONTESTANT_B_eagle-5-1topping the baseline at NAV1.1971(+19.71%).
"The weather outside the cage is public knowledge. How the animals inside handled the storm remains sealed until embargo lift."
- The Market Backdrop: Following two months of steady trending conditions, the first week of September 2026 delivered sharp regime volatility across broader Chinese equity benchmarks.
- The September 02 Breadth Shock (Arena Designation: "Black Wednesday"): Broader Chinese equity benchmarks experienced an abrupt intraday market breadth shock and turnover contraction. Growth-heavy and previously strong market segments suffered sharp localized drawdowns, pushing over 80% of traded equities into localized pullbacks within a 48-hour window.
- The Structural Questions on the Table:
- For quantitative systems whose weights were frozen before July 2026, did prolonged signal persistence hold up when market leadership abruptly rotated?
- How did high-turnover execution handlers (Rabbits) fare against delayed, multi-week inertia strategies (Sloths & Snails) during severe intraday turbulence?
- Did extreme selection intensity (Eagles) protect capital or amplify the drawdown?
⏳ Broadcast Embargo: Episode 09 evaluation data and empirical dispatches unlock on Friday, September 11, 2026.
QuantPits Arena is built on six foundational methodological principles:
- Historical Biography ≠ Arena Record: Past in-sample backtests, former production roles, and historical accolades exist only as biography. They contribute zero weight to Arena standings. Every model enters the common Arena at the same anchor date (2026-07-03) with a normalized base NAV of 1.0000 and CNY 500,000 initial capital.
- High-Resolution Parametric Monkey Null Benchmark: Absolute return alone says very little about signal quality. For each of the 11 portfolio execution policies, an independent colony of 1,000 deterministic pseudo-random monkey portfolios (11,000 total) is evaluated under identical capital and lot constraints.
-
Statistical Interpretation Nuance: A nominal
$p_{\text{upper}} < 0.05$ indicates the result is difficult to reproduce with the random-ranking null. The Arena deliberately does NOT translate this into "Alpha confirmed." - The 28-Animal Execution Zoo: The same underlying model signal is deployed through 28 distinct portfolio behaviors to isolate signal decay, holding inertia, turnover friction, capacity scalability, and directional polarity.
- Real Trading & Finite-Capital Constraints: Fixed initial capital (CNY 500,000), 100-share trading-lot roundings, unaffordable order skipping (preserving real cash drag), and standard transaction costs.
- External Simplicity Reference (The Rock): The Permanent Portfolio is displayed alongside the arena as a simplicity benchmark: "Four buckets. No ranking. No ensemble. No retraining. No comment. Complexity is not free."
| Family | Animals | Specification | Research Hypothesis Tested |
|---|---|---|---|
| Baseline Control | 🤖 Robot | TopK=22, DropN=3, Weekly | Canonical production baseline policy. |
| Cash Lag | 🦥 Sloth 1–4 | 1, 2, 3, 4 Weeks Delay | Delays execution while holding cash; measures signal timing decay. |
| Holding Inertia | 🐌 Snail 1–4 | 1, 2, 3, 4 Weeks Lag | Delays rebalances while holding stale positions; measures inertia friction. |
| Turnover Extremes | 🐢 Turtle 🐇 Rabbit 1–2 |
DropN=1 (Low) DropN=11 / 22 (High) |
Tests whether aggressive deployment extracts alpha or ingests noise. |
| Polarity Inversion | 🐨 Koala | Bottom-22 Worst Stocks | Directional sanity check: inverted ranking should underperform baseline. |
| Rank Geometry | 🦡 Meerkat 10%–90% | 9 Decile Slices (P10–P90) | Tests cross-sectional monotonicity and signal depth across deciles. |
| Concentration | 🦅 Eagle Suite | 5/1, 11/2, 44/6, 66/9, 88/12 | Evaluates concentration risk vs. diversification capacity. |
| Market Breadth | 🐋 Whale Shark 🐉 Taotie |
50% Universe (123 stocks) 100% Full Universe (246) |
Tests survival of signal under broad diversification up to the passive limit. |
QuantPits Arena maintains three distinct tiers of benchmarks answering different research questions:
- CSI 300 Index (SH000300) — External Market Anchor: Reflects external mainland China broad equity market performance and macro context. Not a selection-skill benchmark.
- Taotie — Executable Universe Benchmark: A capital-constrained (CNY 500,000) and round-lot-constrained full-universe portfolio designed to approximate broad exposure with minimal active selection. It is not a theoretical fractional equal-weight index.
- Matched Monkey Distributions — Statistical Selection Null: 11 unique specifications × 1,000 deterministic pseudo-random ranking simulations (11,000 total). Primary statistical reference isolating model ranking signal from portfolio mechanics.
- The Rock (Permanent Portfolio): External 4-bucket simplicity anchor (Equity, Gold, Bonds, Cash) questioning whether model complexity was worth it.
The frontend application (web/) is an institutional, dependency-free single-page research dashboard:
web/
├── index.html # Single-page application shell
├── css/
│ ├── variables.css # Institutional color palettes & design tokens
│ ├── main.css # Typography & layout fundamentals
│ ├── components.css # Tables, cards, compact filter bars, buttons
│ └── views.css # Specific view layouts
└── js/
├── data/
│ └── arena_data.js # Pre-compiled offline tournament payload (168 paths, 11k monkeys)
├── adapter.js # In-memory query engine & statistical aggregator
├── components/
│ ├── charts.js # Apache ECharts wrapper (trajectories, heatmaps, regret curves)
│ └── filters.js # Single-row compact filter toolbar
├── views/
│ ├── landing.js # Platform Exhibition & tournament introduction (/#intro)
│ ├── overview.js # KPI metrics, scatter distributions & standings (/#overview)
│ ├── leaderboard_view.js # Sortable leaderboard & Model × Animal heatmap matrix (/#leaderboard)
│ ├── animals_view.js # 28 Animal containers, dual-scope telemetry & curves (/#animals)
│ ├── contestant_detail.js# Model biography, burial annualized returns & fingerprints (/#contestants)
│ ├── decision_audit.js # Counterfactual decision archaeology & regret analysis (/#decision-audit)
│ ├── methodology.js # Statistical axioms & experimental standards (/#methodology)
│ └── disclaimer.js # Full 8-section legal and research disclosures (/#disclaimer)
└── app.js # Client-side hash router & lifecycle controller
To audit repository integrity and data formatting consistency:
python3 scripts/audit_privacy.pyFor research, educational, and informational purposes only.
Nothing contained herein constitutes investment advice, a recommendation, endorsement, or solicitation to buy or sell any security or financial product.
Results may include live, delayed, simulated, backtested, counterfactual, or shadow-trading performance and are not indicative of future performance. Data are published with an approximately one-week delay. Investing involves risk, including the possible loss of principal.“QuantPits Arena is a research testbed. It studies models, portfolio policies, historical decisions, failures, and occasionally monkeys. It does not tell you what to buy.”
For full disclosures, visit the Research & Legal Disclaimer.
MIT License. Copyright (c) QuantPits Research.