Season — July 2026

Current standings

What you submit

A Python file defining class MyStrategy(Strategy) with one method, on_day, called once per trading day. It receives that day’s data and your current book, and returns target portfolio weights — one number per stock (positive = long, negative = short). The grader normalises your weights to gross leverage 1 and rebalances toward them.

from convexpi.lab import Strategy
import numpy as np

class MyStrategy(Strategy):
    def on_day(self, day, features, prices, portfolio):
        # features: dict of cross-sectional signals, each an array over stocks
        sig = np.nan_to_num(features.get("mom_1m", np.zeros(len(prices))))
        gross = np.abs(sig).sum()
        return sig / gross if gross > 0 else np.zeros(len(prices))

features = dict of per-stock signal arrays · prices = today’s prices · portfolio = your current holdings · return = weights array (len = #stocks). Also gradable in R or Julia — define on_day(day, features, prices, portfolio); the same engine scores every language identically.

How you’re scored

Ranked by Out-of-sample (OOS) Sharpe ratio.

• The market is split into an in-sample window you may fit on and a hidden out-of-sample window you never see while building.
• Your strategy is run over the OOS window, rebalancing weekly toward your target weights.
• Transaction costs are charged on every rebalance (in basis points of turnover).
• Sharpe is the annualised mean/volatility of the resulting daily returns (× √252).
• You are ranked by OOS Sharpe. We also report the overfitting ratio (OOS ÷ in-sample Sharpe) — aim for ≥ 0.7.

The data

A simulated cross-section of stocks over a multi-year daily horizon, with a few planted alpha signals hidden among many noise features. Split into an in-sample window (fit here) and a hidden out-of-sample window (scored here).

features dict of cross-sectional signals (momentum, value, volatility, …) — a few predict returns, most are noise you must filter out.

prices array of each day’s prices, one per stock.

portfolio your current holdings going into the day.

Getting the data you fit on: In the starter notebook, market.features("train") and market.prices("train") give you the in-sample data to fit and validate on; the test window is held out for scoring.

Approaches to try

• Start simple: rank stocks by one signal (e.g. momentum), go long the top and short the bottom.
• Combine a few uncorrelated signals — diversification lifts Sharpe more than tuning one harder.
• Prefer fewer parameters: every knob you fit in-sample is a chance to overfit.
• Compare in-sample vs out-of-sample before submitting — if OOS collapses, simplify.

Just a foothold — the missions and starter notebook go deeper.

Timeline & rules

Get started

1. Open the starter notebook — get the in-sample data, fit a strategy, and check it out-of-sample.
2. Create a free account and submit from the editor.
3. Your OOS Sharpe posts to the leaderboard in under 5 minutes.