CrunchDAO Docs V3
  • Crunch Hub
    • The Crunch-Hub
      • Activity Graphs
  • Competitions
    • Competitions
      • DataCrunch Competition
      • Broad Institute Autoimmune Disease
        • Crunch 1 – Oct 28 to Feb 9 – Predict gene expression
        • Crunch 2 – Nov 18 to Mar 21 – Predicting Unseen Genes
        • Crunch 3 – Dec 9 to Apr 30 – Identifying Gene
        • Full Specifications
        • Lectures
      • ADIA Lab Structural Break Challenge
      • ADIA Lab Causal Discovery
      • ADIA Lab Market Prediction Competition
    • Rallies
      • Mid+One
      • DataCrunch Rally
      • X-Alpha Rally
    • Participate
    • Teams
      • Managing
      • Referendums
      • Leaderboard
      • Rewards
    • Data
    • Code Interface
    • Leaderboard
      • Duplicate Predictions
    • Resources Limit
    • Whitelisted Libraries
    • Known Issues
  • CRUNCH Token practical
    • Release Map
  • Credits
    • Avatar
  • Other
    • Glossary
Powered by GitBook
On this page
  • TL;DR
  • Problem Statement
  • Evaluation
  • Only one Model on the Leaderboard
  • Time constraints
  • Out-of-Range badge
  • Phases and Format
  • Timeline
  • Submission Phase
  • Out-of-Sample Phase
  • Data
  • Building an Attacker with the Mid+One package
  • Some Concepts
  • Example Attackers
  • Attackers FAQ
  • Prizes
  1. Competitions
  2. Rallies

Mid+One

Attacking together, make us stronger.

PreviousRalliesNextDataCrunch Rally

Last updated 2 days ago

Welcome to Mid+One! Dive into the world of martingales and market dynamics. Spot tiny shifts in high-frequency time-series, to predict where prices are heading. It's all about finding that elusive mid-price, one minute into the future.

Ready to attack?

With Mid+One Crunch has found another opportunity: Thousands of banks are consuming mid-market prices for their execution algorithms. The community meta-model will unlock a stream of ongoing rewards with a potential to serve a multi-billion dollar market.

TL;DR

  • Detect small exceptions to the martingale property of a time-series.

  • Determine when a time-series will rise or fall.

  • A "buy and hold" strategy is applied for each prediction over the next 30 time steps.

  • The goal is to maximize profit after accounting for the transaction costs.

  • Only one attacker can be selected for OOS and Reward

  • .

Problem Statement

You will build algorithms that takes one data point at a time and decides whether the average future value 30 steps in advance will be higher or lower than the present value. You only have the time-series, nothing else. Your prediction must be determined only by the past history of the time-series and by patterns you detect therein.

Unlike most forecasting tasks, however, you goal is not to provide a precise prediction 30 steps in the future. Instead you should decide between three possibilities:

  1. The time-series will go up, on average, by at least EPSILON;

  2. The time-series will go down, on average, by at least EPSILON;

  3. Or the average value of the time series will fall between -EPSILON and EPSILON.

EPSILON is the Transaction cost value and is set at 0.0025

Evaluation

For every non-zero prediction, the system initiates a "buy and hold" for 30 data points.

If the prediction is positive we go buy and hold.

If the prediction is negative we go short and hold.

However, a fixed transaction cost (EPSILON) is applied to the profit in both case.

Example

If the value rises by 0.50 over the next 30 periods; the profit will be 0.50 and the net profit would be 0.4975.

Similarly, should the price fall by 0.20 then the net profit would be -0.2025.

Only one Model on the Leaderboard

In the second Rally, you have to choose which model will appear on the leaderboard.

You can still play with 4 different models.

Time constraints

Your tick and predict must run in less then 20ms!

In Mid+One, delivering value to the customer quickly is crucial. Crunch aims to tackle increasingly lower frequencies.

As a result, in the infer function, your code must not take more than 20ms from receiving the message to returning the result. If your average inference time exceeds this limit, your position on the leaderboard will be marked with an "Out-Of-Range" badge.

Out-of-Range badge

Submissions that are too slow but still achieve great results will be rewarded for the Rally.

However, they cannot be deployed in Production or used by Financial Institutions. If you're out of range, there are plenty of ways to optimize your code to meet the 20ms threshold.

Keep in mind that gains in Production will be much higher than during the Rally. OPTIMIZE!

Phases and Format

Timeline

Mid+One is going to evolve into a live Crunch. We went through a first 2 months test phase called "Rally" in order to ensure both problem statement, data and models integrity.

  • Friday Oct 18, 2024, 09:00 AM CET - First rally open

  • Wednesday Dec 18 , 2024, 09:00 AM CET - First Out-of-Sample scoring

  • Wednesday Jan 8, 2025, 09:00 AM CET - Submission re-open - Second Rally

  • Sunday Feb 16, 2025, 11:59 PM CET - Out-of-Sample - Second Rally

  • Live is soon to be announced

Live refers to "in Production" mode where Crunch and the submitted models actively serve real-world end customers.

Submission Phase

Out-of-Sample Phase

Data

# Print the first x_train time-serie's content
for message in x_train[0]: 
    print(message)

# Would print
# ({x: 10303.346153849048})
# ({x: 10303.500000002896})
# ({x: 10303.461538464431})
# ({x: 10303.461538464431})
...
# ({x: 10303.269230772126})
# ({x: 10303.384615387501})
# ({x: 10303.307692310584})

Building an Attacker with the Mid+One package

This package is intended to make life simpler for those participating in Mid+One.

$ pip install --upgrade midone

Some Concepts

  • Attacker is a Python class that consume a univariate sequence of numerical data points (such as stock prices, bond prices, or any time series) x1, x2, …xt and attempts to predict its future movement.

  • Tick is a method from the Attacker class that allows the consumption of incoming data points.

  • Predict is a method from the Attacker class that take a decision base on previous data points.

  • Tick&Predict is a method from the Attacker class that do Tick and Predict in a single function call.

  • Accounting handle tracking and logging the profit and loss (PnL) for decisions made by an Attacker.

Example Attackers

Notebook
Description

A minimalist contest entry notebook

Illustrates use of the Attacker class

Illustrates use of running calculations

Illustrates running regression pattern

Attackers FAQ

Prizes

  • In the first Rally, the top ten performers judged by profit and loss will share $10,000 in proportion to their profit in the out of sample period. (done)

  • In the second Rally, the top 50 performers with positive profit and loss will share $10,000 in proportion to their profit in the out of sample period.

During the the Crunchers are required to submit valid Notebooks or Python files. This submission need to "run" successfully on the Crunch hub in order to receive to be triggered in and receive live data.

Once the start, new data will be run through the models submitted.

In Mid+One, participants are facing univariate time-series called . Crunch's Streams are iterator objects that allow you to traverse all elements of a time-series, one at a time.

Read more in .

Some common questions have already been answered in the .

Quickstarter notebook
Get started quickly with a Quickstarter!
Find the Mid+One package's code here.
attacker.md
FAQ.md
An example is available here.
Mean reversion
Mean reversion attacker
Momentum attacker
Regression Attacker
Learn more how to select your model...
Submission Phase
Out-of-Sample Phase
Out-of-Sample Phase
Streams
Page cover image