Quant Competition Research Update — 2025/2026

Sourced from live top-team writeups, winner interviews, and competition announcements

Last updated: April 2026

1. IMC PROSPERITY 3 — DEEP DIVE (2025 Results + Lessons for Prosperity 4)

Competition Scale

12,000+ teams worldwide, 5 rounds over 15 days
Winner: ~1,433,876 SeaShells (2nd place: Frankfurt Hedgehogs)
9th global / 2nd USA: Alpha Animals — 1,190,077 SeaShells

ROUND-BY-ROUND CONFIRMED WINNING STRATEGIES

Round 1: Rainforest Resin, Kelp, Squid Ink

Product	Winning Approach	Key Detail
Rainforest Resin	Fixed fair value = 10,000. Market make ±2 spread	Buy below 10k, sell above 10k. Also market-take favorable orders.
Kelp	8-period SMA or Wall Mid as fair value. Quote inside spread	"Wall Mid" = (best_bid + best_ask)/2 works as baseline
Squid Ink	Two approaches both worked: (a) z-score mean reversion at >3σ, (b) track Olivia's daily min/max	Squid tripped up many teams — either go all-in or skip it early

2nd place (Frankfurt Hedgehogs) PnL:

Resin: ~39,000 SeaShells/round
Kelp: ~5,000 SeaShells/round
Squid: ~8,000 SeaShells/round (via Olivia pattern detection)

Olivia exploit (Round 1): Bot trader "Olivia" buys exactly at daily lows and sells at daily highs. Track the running daily min/max. When she trades at the extreme, mirror her position before the next reversal.

Round 2: Basket Arbitrage (Picnic Baskets, Croissants, Jams, Djembes)

Basket 1 = 6×Croissants + 3×Jams + 1×Djembe
Basket 2 = 4×Croissants + 2×Jams

What worked:

Compute synthetic fair value from components
Enter when basket deviates beyond fixed threshold (e.g., ±50)
Baskets are mean-reverting relative to synthetic value
Do NOT hedge by trading components — components don't move to match baskets

Key parameter insight: Fixed threshold (e.g., ±50 SeaShells) outperformed dynamic z-score thresholds. Choose parameters from the "stable, flat region" of the performance landscape — not the peak. Peak-optimized parameters overfit to backtest data.

Frankfurt Hedgehogs Result: 40,000–60,000 SeaShells (baskets) + 20,000 SeaShells (constituents) per round

What failed:

Alpha Animals: had bugs in position limit handling that prevented constituent trading → lost significant opportunity
Teams that hedged aggressively with components lost money when components didn't mean-revert

Round 3: Volcanic Rock + 5 Vouchers (Options)

Every top team used Black-Scholes. This is not optional.

Volatility smile strategy (2nd place globally):

Compute implied volatility for each of the 5 strikes from market prices
Plot IV vs. moneyness = ln(S/K)/√T — a parabolic smile appears
Fit a parabola to the smile across strikes
Subtract fitted smile from observed IV for each voucher
Trade the deviations: buy vouchers with IV below fitted curve, sell those above it
Bet on mean reversion of IV deviations back to the smile

Alpha Animals approach:

Pure Black-Scholes fair value per voucher + cross-strike arbitrage
Accidentally held max short Volcanic Rock position due to a bug — which worked spectacularly (rock fell)
Net result: peak 2nd globally that round

What failed:

Volatility surface fitting (full 3D surface) — too complex for 48-hour round timeframe
Delta hedging — calibration issues within position constraints

Round 4: Magnificent Macarons (Cross-Exchange / Conversion Arbitrage)

Core mechanic: Convert between local and foreign exchange accounting for transport fees + tariffs.

Winning approaches:

Volume imbalance exploit: If best_bid_volume > 9 on local exchange → a sell order at best_ask is guaranteed to fill immediately (bots provide that liquidity). Place sell at best ask, guaranteed execution.
Sunlight regime detection: Low sunlight index signals higher production costs → shift from two-way arbitrage to accumulate long, minimize exports
Olivia signal integration: Use her croissant trades as a regime indicator for basket threshold adjustment

What failed:

Linear regression with 99% R² on Macaron prices → classic overfitting. One team lost 82,558 SeaShells in a single round from this bug. If R² is suspiciously high, verify on held-out data immediately.
Sunlight regime modeling: suspected overfitting → one top team abandoned it entirely

Round 5: All Products + Insider Detection

Olivia copy-trading:

Track her win rate over a rolling window (50 trades)
If win rate > 70% on a product → copy her trades as a signal
Alpha Animals: applied to Squid Ink and Croissants → significant P&L

EMA z-score for Squid recovery:

Simple 10-tick rolling window z-score for Squid Ink
Exponential moving average + z-score exploiting "long runs off the price"
More robust than complex forecasting models in the final round

Manual trading challenges:

Rank positions by expected value (probability × payoff), not raw multiplier
Be fee-aware when sizing
Contrarian approach preserves capital but leaves upside unrealized — calibrate risk tolerance to your rank position

PROSPERITY 3 → PROSPERITY 4 STRATEGIC LESSONS

Understand the market structure before building algorithms. Build a custom visualizer (plot order books, mid-prices, spreads over time) before coding any strategy. Teams that did this built deeper intuition and better strategies.
Simplicity beats complexity under time pressure. The 2nd-place team used straightforward first-principles reasoning. The 73rd-place writeup confirmed: replacing complex forecasting with a 10-tick rolling z-score outperformed the complex model.
Parameter stability > parameter optimality. Choose parameters from flat regions of the performance landscape. Peak-optimized parameters overfit.
Bugs cost more than missed strategies. Position limit bugs, unchecked edge cases in trade execution, and uncaught exceptions in single-product logic can cascade and wipe entire rounds. Defensive coding > clever coding.
Don't let a working strategy drift. If a basket strategy that was profitable becomes unprofitable (regime shift), stop it. Don't continue running it hoping it will recover.
Insider detection is always worth implementing. Counterparty data is provided. Track win rates. Olivia appeared in multiple Prosperity editions and was exploitable in both.
Read past-year repos, not just for strategies — for bugs. The most common bugs across all teams: incorrect position limit calculations, handling of partial fills, not resetting rolling windows between rounds.

PROSPERITY 4 SPECIFIC — ANNOUNCED DETAILS (April 2026)

Dates: Tutorial Round March 16 – April 13. Official rounds April 14–30, 2026.
Theme: Outer space
Format: 5 rounds, each with 1 algorithmic + 1 manual challenge
Focus areas announced: Market making, options trading, risk analysis
Preparation: Basic Python + trading concepts (bid/ask, order books, market making, position management)

For Prosperity 4 prep:

Install prosperity3bt — community backtester for P3 data, gives deep familiarity with the framework
Study P3 repos: github.com/TimoDiehm/imc-prosperity-3 and github.com/CarterT27/imc-prosperity-3
Black-Scholes + Greeks will almost certainly appear in Round 3 again. Implement from scratch.

2. WORLDQUANT IQC — 2025 RESULTS + 2026 PREPARATION

2025 IQC Scale

~80,000 participants, 11,000 universities, 142 countries
263,000+ alphas submitted
12 finalist teams (top 0.02%) competed in Singapore, September 2025

2026 IQC Timeline

Stage 1 (Qualifier): March 17 – May 18, 2026
Global Finals: September 2026 (Singapore)
Prize pool: $100,000

ALPHA BUILDING — CONFIRMED WINNING TECHNIQUES

The Fitness Formula (Memorize This)

Fitness = sqrt(abs(Returns) / max(Turnover, 0.125)) * Sharpe

Sharpe must be > 1.25 (Delay-1) or > 2.0 (Delay-0)
Fitness must be > 1.0 (Delay-1) or > 1.3 (Delay-0)
New alpha must have PNL correlation < 0.7 with existing alphas (or Sharpe 10% higher)

Volume Over Perfection

~100 simulations per submittable alpha is typical
One documented winner tested 1,103 alphas, submitted 28 (2.5% pass rate)
Submit every day. Teams that win submit consistently, not in bursts.

The Alpha Patterns That Actually Work (Documented)

1. Price mean reversion (most consistently effective):

rank(-ts_delta(close, 5))

Stocks that dropped most in 5 days tend to rebound. Simple, robust, low turnover.

2. Volume-price divergence:

rank(-correlation(volume, close, 10))

Price up + volume down = weak signal. Negative correlation predicts reversal.

3. Fundamental value ranking:

alpha = ts_rank(cashflow_op/cap, 60);
group_rank(alpha, subindustry)

Operational efficiency vs. peers. Quarterly data = naturally low turnover = better fitness.

4. VWAP reversion:

(vwap - close) / vwap

Compare current price to daily VWAP. Cheap relative to daily average = buy signal.

5. Volume-conditional momentum:

event = volume > adv20;
alpha = (-ts_delta(close, 5));
trade_when(event, alpha, -1)

Only trade mean reversion when volume confirms (above 20-day average volume).

6. News + reversion hybrid: Combine momentum from news coverage with price reversion. Momentum for high-buzz stocks, reversion for low-buzz stocks (conditional logic).

7. Contrarian buzz detection: Short stocks with buzz far above their 60-day average. "Increasing buzz" can signal overheated positions.

Neutralization — The Critical Lever

Market neutralization: Equal long and short weight market-wide. Bets on relative movements only.
Sector/Industry/Subindustry neutralization: Removes sector-level bias. Isolates stock-specific signal.
Neutralizing by subindustry instead of market: reduces raw Sharpe but greatly increases Fitness (the metric that counts for IQC).
Use: group_neutralize(alpha, subindustry) as a default starting point for any fundamental alpha.

Turnover Optimization

Turnover is the enemy. Every point of unnecessary turnover kills Fitness.
Use decay_linear(alpha, N) to smooth signals and reduce daily turnover
Fundamental data (quarterly) inherently has low turnover — use it heavily
ts_rank(signal, 60) instead of raw signal smooths noise and reduces flipping

Decorrelation = Your Competitive Advantage as a Team

IQC scores teams on a basket of alphas. A portfolio of 10 uncorrelated alphas with Sharpe 1.5 beats 10 correlated alphas with Sharpe 2.5.

Decorrelation checklist:

Mix data sources: price/volume + fundamental + news/sentiment
Mix signal types: momentum + reversion (they tend to be decorrelated)
Mix time horizons: 1–5 day signals vs. 20–60 day signals
Check PNL correlation before submitting. If > 0.7 with existing, only submit if Sharpe is 10%+ higher.

The "101 Formulaic Alphas" Paper

A single academic paper by Kakushadze & Tulchinsky contains 101 directly testable alpha formulas. Systematically implementing all 101 will yield multiple qualifying alphas. This is the single highest-ROI resource for IQC.

Search: "101 formulaic alphas" on Google Scholar or SSRN

3. JANE STREET — WHAT ACTUALLY MATTERS IN 2025/2026

Jane Street does not run a traditional "win or lose" competition. The programs to target are:

Electronic Trading Challenge (ETC)

Format: Day-long programming contest, teams trade multiple simulated instruments (bonds, stocks, ETFs, ADRs) against each other via a live exchange.

Confirmed winning strategies from documented participants:

Bonds:

Fair value = fixed at 1,000. Buy any bond offered below 1,000. Sell at 1,001.
Simple, nearly risk-free. Dominated early competition until all teams discovered it.

ADRs:

Do NOT use book offers to estimate fair value — too noisy.
Use the most recent TRADE PRICE of the liquid ADR as the valuation benchmark.
Apply ±10 USD margins around that trade price.
After every 10 units, convert between liquid and illiquid form to hedge.
Result: ~$2,000 per round profit after this fix (vs. near-zero before).

ETFs:

Value the ETF as: bond_weight × bond_price + stock_weights × stock_prices.
Update stock valuations using most recent trade prices (not book).
Apply ±30 USD margins due to higher inherent volatility.
Result: ~$7,000 per round — the highest-margin opportunity documented.

Infrastructure tips:

Use Python. Prioritize readability and debuggability over performance.
Use randomized order placement intervals, not constant — prevents "broken pipe" errors.
Add 0.01-second delays between consecutive trades to prevent server rejection.
Set up Git aliases before the competition starts. Every second counts.
Test locally against your own simulated exchange before going live.

Key insight: In every instrument category, the pattern is the same — use TRADE prices, not book prices, to estimate fair value. Book offers are often stale or manipulated by other bots.

Monthly Puzzles (Highest Sustained ROI)

Jane Street recruiters explicitly monitor the monthly puzzle leaderboard.
Consistent puzzle solving = cold recruiter outreach from Jane Street.
Archive available: janestreet.com/puzzles/archive/
Difficulty level: requires combinatorics, probability, and optimization depth. Not trivial.

Kaggle: Jane Street Market Prediction

Jane Street has hosted Kaggle competitions (market prediction, real-time forecasting)
Top approaches: neural networks (autoencoders + MLP), ensemble methods
Core challenge: the feature-response relationship constantly changes — avoiding temporal overfitting is the #1 technical hurdle
Fold strategy: split by day (not randomly). Mutual exclusivity between folds prevents day-information leakage.
Custom imputer + XGBoost performed strongly on Jane Street's 2024 competition

4. CITADEL DATATHON — CONFIRMED WINNING APPROACH

What Judges Actually Score

The Citadel datathon is not a modeling competition — it is an insight competition with a modeling component. Judges evaluate:

Quality and novelty of the research question
Quality of analysis supporting it
Clarity and persuasiveness of the presentation
Robustness of methodology

The Winning Formula (From 2021 PhD Datathon Winner + Documented Teams)

Step 1: Research question selection (most impactful step)

Choose a question with clear human or societal impact
The question must be answerable with the tools your team knows well
Narrow down from a broad literature survey — start with wide reading, then focus
"Find a research question which is impactful and can be answered well using the tools you know."

Step 2: Data strategy

Actively supplement the provided dataset with higher-quality external data
Source from government databases, academic repositories, open data portals
Document your data sourcing — judges want to see rigor
When you hit a technical obstacle (unfamiliar file format, API rate limit), solve it — don't switch topics

Step 3: Analysis

EDA first. Always. Before any modeling.
Feature engineering: remove highly correlated features, scale appropriately, handle missing values explicitly
Ensemble methods outperform single models consistently in datathon settings
Gradient boosting (XGBoost, LightGBM) + SHAP analysis = judges see what your model learned
State your assumptions explicitly. One wrong assumption discovered during Q&A destroys credibility.

Step 4: Presentation (this is 50% of winning)

Every chart must be self-contained: caption should convey the full message without reading the text
Every team member must present a section — judges notice when one person does all the talking
Structure: context → key insight → evidence → implications → limitations → next steps
Acknowledge limitations before judges ask — shows intellectual honesty and earns trust
Practice the pitch out loud, multiple times, under time pressure
End with concrete, actionable next steps — "future work" should be specific, not vague

Step 5: Q&A defense

Prepare for: "Why didn't you do X?" — have answers ready for obvious alternatives
Prepare for: "How robust is this to Y?" — test at least 2 robustness checks
Prepare for: "What would change your conclusion?" — show you understand your own assumptions

Technical Stack That Works

Python: pandas, scikit-learn, XGBoost/LightGBM, matplotlib/seaborn/plotly
SHAP for model interpretability (judges love seeing this)
statsmodels for hypothesis testing (adds rigor)
Clean Jupyter notebooks you can present directly

5. MACHINE LEARNING vs. MATHEMATICAL APPROACHES — WHAT ACTUALLY WINS IN 2025

This is the most important meta-question. Here is what the evidence shows:

The Short Answer: Hybrid Wins, Pure ML Often Fails

Evidence from IMC Prosperity 3:

Pure ML models (linear regression with R²=99%) caused catastrophic losses
Simple mathematical models (fixed threshold, z-score, Black-Scholes) placed top 10 globally
The 2nd-place team prioritized "deep structural understanding" over ML sophistication
ML was used for: parameter tuning (grid search), detecting regime patterns

Evidence from WorldQuant IQC:

Simple algebraic alpha expressions outperform complex ML models consistently
rank(-ts_delta(close, 5)) — literally 1 line — is among the top-performing alpha patterns
The platform constraints (expression-based operators) discourage raw ML
ML is useful for: hypothesis generation, feature importance analysis, identifying which fundamental factors matter

Evidence from Jane Street Kaggle:

Neural networks (autoencoders + MLP) dominated the leaderboard
BUT: the challenge is specifically framed as a prediction task with tabular features
The hardest problem is temporal overfitting, not model capacity
ML wins when: (a) the task has a clear prediction target, (b) features are available, (c) there is sufficient non-overlapping training data

Evidence from Citadel Datathon:

Gradient boosting + SHAP is the standard winning stack
Complex deep learning adds no advantage over interpretable ensemble methods for judges
The bottleneck is insight quality + presentation, not model sophistication

Framework: When to Use What

Situation	Use Math/Statistics	Use ML
Real-time algo trading competition (Prosperity)	Primary tool	Parameter tuning only
Alpha expression platforms (WorldQuant)	Primary tool	Feature discovery
Prediction competitions (Kaggle/Jane Street)	Regularization, feature engineering	Primary model
Datathons (Citadel)	Hypothesis testing, EDA	Gradient boosting + SHAP
Options/derivatives (any competition)	Required (Black-Scholes, Greeks)	Never

The Most Dangerous Failure Mode: ML Overfitting

Every winning writeup from 2025 mentions this in some form:

99% R² that collapses live → the single most expensive mistake in algo competitions
Neural networks trained on price sequences without proper time-based cross-validation → always fails out-of-sample
Hyperparameter optimization on the full dataset → silent overfitting

The fix: Time-based splits (train on earlier time, validate on later time). No exceptions. Never shuffle time-series data for CV.

What Quant Firms Actually Want to See

From Jane Street, Citadel, and IMC recruiter signals:

Evidence of thinking about WHY a strategy works, not just that it worked on backtest
Ability to identify and quantify edge decay
Robust risk management reasoning (position limits, VaR thinking, drawdown awareness)
Clean, readable, well-commented code that could be reviewed in 5 minutes
Understanding of market microstructure (bid-ask spread, slippage, adverse selection)

6. GENERAL QUANT COMPETITION STRATEGY — PRINCIPLES THAT WIN

Before the Competition Starts

Study prior-year solutions. Not to copy — to understand the product space, common traps, and what the organizers reward. GitHub has repos for every major competition.
Build a backtest environment first. Before day 1, have a framework where you can test ideas in 5 minutes, not 2 hours.
Assign roles explicitly. One person owns: (a) backtesting pipeline, (b) live strategy deployment, (c) data analysis/visualization. Overlap is fine but unclear ownership kills speed.
Prepare your debugging toolkit. Have logging, position tracking, and PnL attribution built before round 1 starts.

During the Competition

EDA before algorithms. Plot everything. Histogram of prices, autocorrelations, cross-correlations between products, order book depth visualization. The fastest path to good strategies is pattern recognition, not model building.
Deploy a simple baseline immediately. A market maker on the most stable product is worth deploying in round 1, hour 1. It earns while you develop more complex strategies.
Timebox your complexity. If you haven't implemented and tested something in 4 hours, simplify it. The 2nd-place Prosperity 3 team: simple strategies, thoroughly validated.
Monitor live PnL in real time. Attribute P&L to each product and each strategy component. If one component is losing, understand why before the round ends.
Manual challenges are underweighted by most teams. Expected-value maximization for the manual trading rounds is pure math — no coding required and often worth as much as complex algorithm rounds.

After Each Round

Debrief: what worked, what didn't, why?
Read competitor discussions (Discord, GitHub) — in Prosperity 3, the community was active and top teams shared partial hints
Update your fair value models with new round data
For WorldQuant: submit 5+ new alphas based on learnings from the previous day

The Mental Game

Accept calculated risk proportional to your rank position. If you're in top 10, protect. If you're at rank 500 and need top 25 for prizes, you need alpha — take strategic risk.
Don't abandon working strategies. Add on top of them. Only kill a strategy when it's clearly losing money post-regime-change.
Bugs are your biggest competitor. More teams fall from technical errors than from bad strategy theory.

KEY RESOURCES

Resource	URL	Value
Frankfurt Hedgehogs P3 (2nd global)	github.com/TimoDiehm/imc-prosperity-3	Code + writeup
Alpha Animals P3 (9th global, 2nd USA)	github.com/CarterT27/imc-prosperity-3	Code + round writeups
P3 community backtester	github.com/jmerle/imc-prosperity-3-backtester	Essential tool
P3 visualizer	jmerle.github.io/imc-prosperity-3-visualizer/	Order book viz
WorldQuant BRAIN alpha examples	github.com/alexisdpc/WorldQuant-alpha-trading	Formula library
WorldQuant IQC 2026	worldquant.com/brain/iqc/	Official
IMC Prosperity 4	prosperity.imc.com	Official, starts Apr 14
"101 Formulaic Alphas"	Google Scholar / SSRN	IQC alpha bible
Jane Street puzzles archive	janestreet.com/puzzles/archive/	Monthly practice
P3 Martin Oravec writeup (73rd)	medium.com/@oravec.martin01/imc-prosperity-3	Failure lessons
P3 Sam Bennett writeup (74th)	medium.com/@samjgbennett/trading-triumphs	UK top 4 writeup

Compiled from: Alpha Animals GitHub writeup, Frankfurt Hedgehogs GitHub, Ben Hubsch Jane Street ETC writeup, David Veitch Citadel datathon writeup, WorldQuant IQC official announcements, James Glazar IQC project notes, WorldQuant BRAIN documentation, and multiple 2025 competition participant Medium posts.