Bond ETF arbitrage with Machine Learning
Log in to collectAcademic paper
Julio A. Crego; Jens Soerlie Kvaerner; Åvald Åslaugson Sommervoll; Dag Einar Sommervoll; Niek Stevens
- NLTilburg University
- NOUniversity of Oslo
- ?Digital infrastructure and security
- NONorwegian University of Life Sciences
- ?NTNU Business School
- ?Norwegian University of Life Sciences (NMBU)
- Capital University
- ?SIG-i Capital AG
Strategy in a nutshell
1. Target Selection
The strategy begins by identifying ETFs with potential mispricing. Each trading day, ETFs traded consistently over the past 500 days are considered candidate targets. These candidates are ranked based on their premium relative to NAV, and one ETF among the top three is randomly selected to short. This selection method ensures diversification and focuses on ETFs whose price deviates from fundamentals.
2. Mimicking Portfolio Construction
Once a target is selected, a mimicking portfolio of up to eight ETFs is constructed using a genetic algorithm and nonnegative least squares (NNLS). This portfolio consists of long positions only, with carefully constrained weights to minimize transaction costs and maintain liquidity. Daily profits are calculated
Economic rationale
The strategy’s core idea is to create a liquid mimicking portfolio of ETFs that substitutes for the underlying securities in the NAV. By combining a genetic algorithm with nonnegative least squares (NNLS), the method identifies ETFs whose combined returns replicate a target ETF, allowing arbitrageurs to exploit mispricing efficiently. Key requirements include liquidity, low-cost tradability, positive weights, and a high explanatory R² in regression. Constraints—