Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Numin: Weighted-Majority Ensembles for Intraday Trading

Aniruddha Mukherjee, Rekha Singhal, Gautam Shroff

TL;DR

Numin addresses intraday equity forecasting by aggregating predictions from multiple data scientists through a weighted-majority ensemble. It situates the approach between Numerai-style meta-models and real-time trading, employing dynamic EMA-based weight updates guided by two metrics: accuracy and a trading-utility proxy. The study demonstrates that utility-based weighting, particularly over short windows, can improve practical profitability while maintaining competitive accuracy, even when individual models underperform. Overall, the results suggest that WMA ensembles offer a viable, adaptive framework for near real-time model combination and for rewarding contributors in a Numerai-inspired setting.

Abstract

We consider the application of machine learning models for short-term intra-day trading in equities. We envisage a scenario wherein machine learning models are submitted by independent data scientists to predict discretised ten-candle returns every five minutes, in response to five-minute candlestick data provided to them in near real-time. An ensemble model combines these multiple models via a weighted-majority algorithm. The weights of each model are dynamically updated based on the performance of each model, and can also be used to reward model owners. Each model's performance is evaluated according to two different metrics over a recent time window: In addition to accuracy, we also consider a `utility' metric that is a proxy for a model's potential profitability under a particular trading strategy. We present experimental results on real intra-day data that show that our weighted-majority ensemble techniques show improved accuracy as well as utility over any of the individual models, especially using the utility metric to dynamically re-weight models over shorter time-windows.

Numin: Weighted-Majority Ensembles for Intraday Trading

TL;DR

Numin addresses intraday equity forecasting by aggregating predictions from multiple data scientists through a weighted-majority ensemble. It situates the approach between Numerai-style meta-models and real-time trading, employing dynamic EMA-based weight updates guided by two metrics: accuracy and a trading-utility proxy. The study demonstrates that utility-based weighting, particularly over short windows, can improve practical profitability while maintaining competitive accuracy, even when individual models underperform. Overall, the results suggest that WMA ensembles offer a viable, adaptive framework for near real-time model combination and for rewarding contributors in a Numerai-inspired setting.

Abstract

We consider the application of machine learning models for short-term intra-day trading in equities. We envisage a scenario wherein machine learning models are submitted by independent data scientists to predict discretised ten-candle returns every five minutes, in response to five-minute candlestick data provided to them in near real-time. An ensemble model combines these multiple models via a weighted-majority algorithm. The weights of each model are dynamically updated based on the performance of each model, and can also be used to reward model owners. Each model's performance is evaluated according to two different metrics over a recent time window: In addition to accuracy, we also consider a `utility' metric that is a proxy for a model's potential profitability under a particular trading strategy. We present experimental results on real intra-day data that show that our weighted-majority ensemble techniques show improved accuracy as well as utility over any of the individual models, especially using the utility metric to dynamically re-weight models over shorter time-windows.

Paper Structure

This paper contains 25 sections, 3 equations, 3 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Ensemble Learning
  4. Numerai
  5. Methods: Algorithms & Models
  6. Weighted Majority Algorithm (WMA)
  7. Predictive Models
  8. Categorical Convolutional Neural Network (CNN)
  9. Convolutional Neural Network (CNN) Regressor
  10. Long Short-Term Memory (LSTM) Categorical
  11. Long Short-Term Memory (LSTM) Regression
  12. Continual Multi-Layer Perceptron (MLP)
  13. Mixture of Experts (MoE)
  14. AutoEncoder Test-Time Adapter
  15. K-Means Clustering
  16. ...and 10 more sections

Figures (3)

  • Figure 1: Accuracies of the 8 models along with an "AVG_MODEL" and the two WMA methods over 3 different window sizes.
  • Figure 2: Utility of the 8 models along with an "AVG_MODEL" and the two WMA methods over 3 different window sizes.
  • Figure 3: Temporal evolution of normalized weights for eight models using Exponential Moving Average (EMA), a (5,10) window, and $\phi$ set to model accuracy, observed on May 8th yielding a 0.18 Utility.