BigShort
A five-hour machine learning trading sprint. Build a model, survive the crash, then trade your own signal.
How the five hours play out
Calm, then chaos, a breather, then a new game. Each block is one hour; organizers may shift timings slightly on the day.
Kickoff
Get your market
Log in. Your team gets its own stock and its own slice of time to predict. No two teams have the same data, so copying another team's code gets you nowhere.
R1 · 1 h
Foundation
Calm market. Clean the data, engineer features, train a model that says UP (1) or DOWN (0) for each of your 50 prediction rows.
Flash crash
The shock
Mid-event, the market you were studying collapses. Volatility spikes, prices fall, and your Round 1 model is probably broken. New crash data unlocks.
R2 · 1 h
Signal Hunt
Same task, new regime. Retrain, add volatility filters, shorten your lookback. This round tests whether your model adapts.
Break · 1 h
Cooldown
No uploads during the break. The Round 3 trading desk opens when it ends.
R3 · 1 h
Alpha
Stop predicting direction and start trading. For each row, choose a position from -1 (full short) to +1 (full long). Your book is backtested with trading costs.
Finale · 1 h
Presentation
Presentations and evaluation. Final standings are published on the leaderboard once evaluation is done.
What you upload
One CSV per upload. Exactly 50 rows matching the round prediction window (R1: 801–850, R2/R3: 851–900).
Rounds 1 & 2: predictions
Round 1 timestamps: 801–850 | Round 2 timestamps: 851–900
timestamp,prediction,confidence 801,1,0.82 802,0,0.63 803,1,0.55 ...
- -prediction must be 0 or 1
- -confidence is optional, between 0.0 and 1.0
- -exactly 50 rows, timestamps 801 to 850
Round 3: signals
Round 3 timestamps: 851–900
timestamp,signal 851,0.85 852,-0.40 853,0.00 ...
- -signal between -1.0 (full short) and +1.0 (full long)
- -0.0 means flat, no position
- -exactly 50 rows, timestamps 851 to 900
How you're scored
Every round is scored out of 100.
Rounds 1 and 2
Scored on how well your predictions beat a coin flip, using balanced accuracy and the Matthews correlation coefficient. Guessing, or calling UP every time, lands at roughly 17, so that is the baseline to beat.
Round 3
Your positions are backtested with trading costs. Risk-adjusted returns are rewarded; deep drawdowns and trading too often both cost you.
Rules that matter on the day
Read these once, then forget them.
- 0110 submissions in total, shared across all three rounds. Plan how you spend them.// RULE_CONSTRAINT_01
- 02One upload per 60 seconds per team.// RULE_CONSTRAINT_02
- 03Your public score shows as soon as an upload is accepted. It is based on the first 25 of your 50 rows. Leaderboards during the event use public scores; the final ranking uses all 50 rows.// RULE_CONSTRAINT_03
- 04One login per team, given to you by the organizers. Everyone on the team uses it.// RULE_CONSTRAINT_04
Field guide
Terms you will hear a lot.
- Long and short
- Long means buying because you expect the price to rise. Short means selling borrowed shares and buying them back later, profiting if the price falls.
- Read more on CFI (opens in a new tab)
- Alpha
- The excess return that comes from skill, after accounting for general market movement.
- Read more on CFI (opens in a new tab)
- Backtesting
- Testing a strategy on historical data to see how it would have performed. The event scorer works as an automated backtest.
- Read more on CFI (opens in a new tab)
- Overfitting
- A model that memorises its training data instead of learning rules that generalise. Near-perfect accuracy in your own tests is a warning sign.
- Read more on CFI (opens in a new tab)
- Transaction costs and turnover
- Every trade has a cost. A strategy that changes position too often can look profitable until those costs are subtracted.
- Read more on CFI (opens in a new tab)
- Lookahead bias
- Accidentally using information that would not have been available at decision time. It makes results look better than they will be.
- Read more on CFI (opens in a new tab)