How AlphaNova Competitions Work: Format, Evaluation and Prizes

AlphaNova runs quantitative signal forecasting competitions on factor-hedged residuals. You submit a model that ranks assets by expected future returns. We evaluate it out-of-sample, admit it only if it survives a strict quality filter, and pay cash prizes to the top performers.

This page is the complete reference: what you submit, how it is scored, how prize pools are built, and when money changes hands.

If you are new to the field itself, start with How to Become a Quant for context on how signal research careers and competitions fit together.

At a glance

What you forecast: factor-hedged residuals
What you submit: a Python Predictor class
How it is scored: out-of-sample Sharpe ratio, walk-forward
New competition opens: the 1st and 15th of every month
Submission window: two weeks
Live scoring window: one month
Prize pool: up to USD2,500 + carry-over per cycle
Entry fee: none
IP: you retain full ownership

The premise: forecast the residual, not the asset

Most forecasting competitions ask you to predict raw returns. The problem is that raw returns are dominated by a handful of common factors — market beta, value, momentum, size. A model that predicts raw returns well may simply be repackaging factor exposure that everyone already knows about.

AlphaNova asks a harder and more useful question: can you forecast what is left over after factor exposure is removed?

Those leftovers are the residuals. Factor-hedged residuals strip out the returns explained by known risk factors, leaving only the idiosyncratic component. If your model can rank assets by their expected residual returns, you have found something that is, by construction, not just a repackaged factor bet. That is what practitioners mean by genuine alpha.

This design choice shapes everything else about the competition. Because residuals are noisy and low signal-to-noise, evaluation has to be rigorous. Because genuine alpha is rare, admission has to be strict. And because strict admission is the scarce resource, the prize pool is tied directly to it.

What you submit

You submit a Python Predictor class. At its core, the class implements a prediction method that takes a point-in-time view of the universe and returns a ranking of assets by expected future residual return.

The interface is deliberately minimal. AlphaNova handles the data pipeline, the factor neutralisation, the train/test discipline and the scoring. You focus on the signal.

Submissions are point-in-time and walk-forward. Your model only ever sees data that would have been available at the moment the prediction is made. There is no opportunity to leak future information into training, and no reward for doing so — a model that depends on lookahead will fail out-of-sample.

How submissions are evaluated

Evaluation happens on data your model has never seen. Two principles govern it:

Walk-forward. Rather than a single train/test split, evaluation proceeds through time. The model is trained on a window, predicts the next period, and the window rolls forward. This mirrors how a signal would actually be deployed and prevents the model from being tuned to one particular market regime.

Out-of-sample. Every prediction that counts toward your score is made on data withheld from training. In-sample fit is irrelevant to your result.

The scoring metric is the Sharpe ratio of your predictions' realised performance over the live window. Sharpe is used because it rewards consistency rather than raw magnitude — a signal that is right often and wrong mildly beats one that is occasionally spectacular and frequently wrong.

How signals are admitted — and why it matters for the prize pool

Scoring produces a leaderboard. Admission is a separate question.

A model can score well and still be rejected, for two reasons:

Correlation with existing signals. If your model reproduces alpha that AlphaNova has already admitted, it adds no new information. Admitted signals must be sufficiently uncorrelated with what is already in the book.
Overfitting. A model that performs well in-sample but degrades out-of-sample is filtered out. Overfit detection is applied before admission.

Admission uses a greedy quality selection process. Signals are considered in quality order, and each is admitted only if it clears the quality bar and adds distinct information to the existing set. The result is a portfolio of genuinely independent, genuinely predictive signals rather than a pile of correlated ones.

This is where the competition design becomes economically interesting.

A competition's prize pool grows only when genuine, distinct alpha is found. Each cycle starts with a base pot of up to USD2,500. Every quality signal the cycle admits increases the pot. If a cycle admits nothing, the pot does not grow — and unused amounts carry over into future cycles.

The incentive alignment is direct: you are rewarded for finding alpha that the collective book does not already contain, not for finding alpha that duplicates it.

You can model this yourself with the pot calculator on the competition page.

The competition cycle

Competitions run on a fixed, rolling schedule. There is always an active competition.

New competitions open on the 1st and 15th of every month.
Submissions stay open for two weeks. When one competition closes, the next opens immediately.
Closed competitions enter a one-month live scoring window. The window ends exactly one month later, on the same calendar day — a competition that closes on the 1st settles on the 1st; one that closes on the 15th settles on the 15th.
Prizes are paid at settlement, in cash or stablecoin.

Because the two-week submission window and the one-month live window overlap, multiple competitions are in different stages at any given time: one accepting submissions, several in live scoring, and recently completed ones settling.

Prizes and payment

The standard base pool of USD2,500 is split across the top three finishers:

🥇 1st place — USD1,500 (60% of the base pool)
🥈 2nd place — USD625 (25% of the base pool)
🥉 3rd place — USD375 (15% of the base pool)

Where a cycle's pot exceeds the base amount through carry-over and admitted signals, prizes scale with the pot.

Prizes are paid at settlement in cash or stablecoin. There is no staking requirement, no token exposure, no volatility risk and no lock-up period. What you win is what you receive, when the live window closes.

Seasons and formats

Biweekly competitions are grouped into seasons of five contests. Season 1 comprises Biweekly #1 through #5, running on the fixed 1st-and-15th cadence.

Alongside the biweekly series, AlphaNova runs larger flagship competitions with substantially bigger prize pools. Competition 5 carries a USD36,000 pool and runs on the same signal-forecasting foundation as the biweekly series.

What you keep

Two things matter more than most participants expect.

Entry is free. There is no fee to compete in any AlphaNova competition.

You retain full intellectual property. Your model, your code and your research are yours. Competing does not transfer ownership of your work.

Frequently Asked Questions

When do new competitions open?
New competitions open on the 1st and 15th of every month. When one competition closes, the next opens immediately — so there is always an active competition.

How long is the live scoring window?
Exactly one month. A competition that closes on the 1st settles on the 1st of the following month; one that closes on the 15th settles on the 15th of the following month.

How are prizes paid?
At settlement, in cash or stablecoin. There is no staking requirement, no token volatility, and no lock-up period.

Is there an entry fee?
No. Competitions are free to enter, and participants retain full intellectual property over their submissions.

How are winners chosen?
Signals are evaluated walk-forward and out-of-sample using the Sharpe ratio. A greedy quality selection process admits only uncorrelated, overfit-filtered signals — so prize pools grow only when genuine, distinct alpha is found.

Can I enter more than one competition at a time?
Yes. Biweekly competitions overlap by design — a submission in one cycle does not prevent you from submitting in another. Many participants maintain an evolving signal pipeline that they refine across cycles.

Why residual forecasting instead of raw returns?
Raw returns are dominated by common factor exposure. Forecasting residuals isolates the idiosyncratic component, which is where genuine alpha lives and where a model can add information that is not already widely known.

What happens if my model scores well but is not admitted?
You may still place on the leaderboard, but admission to the signal book — and the prize pool growth that follows — requires clearing the correlation and overfitting filters. A high-scoring model that duplicates existing signals does not grow the pot.

Do I need to be an experienced quant to compete?
No. The submission interface is a single Python class, and the platform handles data, neutralisation and evaluation. That said, the quality bar is high — see How to Become a Quant for the research skills that tend to matter.

Getting started

Pick the currently open competition, write a Predictor, and submit before the two-week window closes. If you want to see what the standard looks like before you start, read Meet the Minds behind the Models — an interview with one of our early prizewinners.

Join the competition here.