The HFT firm Wunder Fund's hosting platform has launched the public ML competition Alpha Connectome: participants predict indicators of future price movement based on the order books of two related instruments. The competition started on September 11, 2026, submissions close on November 15, 2026, and the prize pool is $13,600 USDT for the top 8 on the private leaderboard. To win a prize, it is not enough for the top 8 to look good on the public leaderboard: they face code verification and a final recalculation of results on a hidden dataset.
What happened
The Wunder Fund platform opened registration for the Alpha Connectome ML competition (internal code wnn33, status active). Participants receive 112 features: order books with 11 levels of bid/ask prices and volumes for two related instruments i0 and i1, trade history, and 8 additional features. The task is to predict two indicators, t0 and t1, of the future price movement of instrument i0, while the organizers do not disclose the exact definition of these indicators. The data is provided in Parquet: train — 10,607 sequences of 20,000 rows each, totaling 212.14 million rows; validation — 1,873 sequences, approximately 37.46 million rows. Results are evaluated using the Global Weighted Pearson metric with clipping of targets and predictions to the range [-2, 2], averaged over t0 and t1. The limit is 5 submissions per day; the top 8 on the private leaderboard within 7 days after the preliminary results must undergo verification: inference and training code with fixed seeds, a 2–8 page report, and a 30–60 minute review call, after which the final result is recalculated on a hidden dataset. Prize pool breakdown: $5,000 for first place, $2,500 for second, $1,700 for third.
Context
Wunder Fund is an HFT firm's hosting platform that, according to a Telegram post, has been operating since 2014 and has a daily turnover of over $10 billion. Alpha Connectome is the third round of its competition series following Wunder Challenge and Predictorium, with the $13,600 USDT prize pool remaining unchanged from round to round. The problem itself — predicting price movement from the limit order book — is a classic microstructure task with low research novelty; the cross-instrument part is more interesting: what exactly can a model extract from the related instrument i1 and the eight additional features, about which there are almost no details in the sources. A key feature of this round is the anti-overfitting mechanism: hidden targets t0/t1 and final rescoring on a hidden dataset test whether the model finds a real dependency in the order book rather than fitting to a known metric formula.
Why this matters for the industry
Open code competitions have become a hiring pipeline for quants at HFT firms: the $13,600 prize pool is inherently less than the annual salary of even a strong junior, but the company gets a funnel of participants who research its live task on real data for free. Winner verification — code with fixed seeds, a report, and a review call instead of a simple solution checkout — is a direct response to the main failure modes of public leaderboards: fitting to the metric formula and overfitting to a known target. If the format of "hidden targets, rescoring on a hidden dataset, and code audit" is established in the third round, it could set an industry standard for honesty in benchmarks and hiring in the quant segment. After the deadline, the verified reports of the top 8 will form a mini-corpus of validated and reproducible approaches to order book prediction — essentially an open textbook on validation under hidden targets.
Why this matters for users
For readers, this is a free chance to work with order books and trade history firsthand: registration is open on the official website, Parquet data is already available, and the public leaderboard provides quick feedback on each submission. In the first few days, it is realistic to build a Parquet reading pipeline, a basic correlation baseline, and the first submission, after which all the value lies in the discipline of the daily attempt limit and local validation instead of blindly trusting the leaderboard. Winners are promised prizes from $600 to $5,000 USDT and, according to the authors of the Telegram post, an invitation to an interview at Wunder Fund. Even without making the top 8, there remains a ready-made pet project with an open leaderboard, and after the preliminary results — finalist reports as educational material.
What is still unknown / limitations
The exact definition of targets t0/t1 has not been disclosed: this prevents fitting to the metric formula, but effectively turns the task into blind target identification and reduces reproducibility and targeted feature engineering. There are almost no details in the available materials about the eight additional features and what exactly is taken from the related instrument i1. Global Weighted Pearson with clipping is a statistically noisy correlation metric on financial data: with validation of 1,873 sequences, the gap between adjacent places may lie within the statistical error. The size and composition of the hidden dataset for final rescoring are not described. Information about Wunder Fund and the promise of an interview at the end of the round are taken from a Telegram post and have no independent confirmation in the available materials.
Sources
- Alpha Connectome ML Competition — official competition page on the Wunder Fund platform
- Prizes and verification — Alpha Connectome documentation: prize pool, breakdown by place, top-8 verification
- Data overview — Alpha Connectome documentation: 112 features, L11 order books, targets t0/t1, WP metric, data sizes
Author
Look at AI, editorial team