Dynamic Asset Allocation Using Asset-Specific Regime Forecasts: A comparison of Discrete and Continuous Jump Models with XGBoost
Translated title
Dynamisk aktivallokering baseret på aktivspecifikke regimeforudsigelser: En sammenligning af diskrete og kontinuerte jump-modeller med XGBoost
Author
Shamoun, Peter Odisho Malko
Term
4. term
Education
Publication year
2026
Submitted on
2026-10-30
Abstract
This thesis investigates whether forecasting market regimes for individual assets can improve risk-adjusted returns in dynamic asset allocation. Market regimes here mean periods when an asset is typically in a positive (bullish) or negative (bearish) phase. The study compares two statistical methods for identifying such regimes: a Discrete Statistical Jump Model (JM) and a Continuous Jump Model (CJM). For each of 9 assets – spanning equities, fixed income, and commodities – the models are used to label historical periods as bullish or bearish. These historical regimes then serve as training data for XGBoost models, a modern machine learning technique, which aim to forecast the regime for the next trading day. The empirical analysis uses daily total-return data and evaluates performance out of sample from 2008 to 2025. First, the forecasts are tested in asset-specific timing strategies, which decide when to be invested or hold cash. Next, the regime forecasts are incorporated into three types of portfolios: equal-weight (EQ), minimum-variance (MinVar), and mean-variance (MV). These strategies are compared with conventional benchmarks, with transaction costs included. Both jump models succeed in identifying regimes with clearly different return and volatility patterns, but the XGBoost forecasts do not consistently improve the performance of the individual timing strategies. The portfolio results are mixed as well. The largest improvements appear in minimum-variance portfolios, where JM-XGB and CJM-XGB achieve Sharpe ratios of 1.185 and 1.285, respectively, compared with 0.877 for the conventional benchmark, and they also experience substantially smaller maximum drawdowns. An analysis of portfolio weights shows that these strategies shifted considerable amounts into cash during the 2008 financial crisis and the 2022 market decline. However, paired stationary bootstrap tests find no statistically significant differences in Sharpe ratios across the eight portfolios that are compared. Overall, the findings suggest that asset-specific regime forecasts can improve risk-adjusted performance under certain allocation rules, but neither the discrete nor the continuous approach consistently outperforms traditional portfolio strategies.
Denne afhandling undersøger, om man kan opnå bedre risikojusteret afkast i dynamisk aktivallokering ved at forudsige markedsregimer for hver enkelt aktivklasse. Markedsregimer betyder her perioder, hvor et aktiv typisk er i en positiv (bullish) eller negativ (bearish) fase. Afhandlingen sammenligner to statistiske metoder til at identificere sådanne regimer: en diskret statistisk jumpmodel (JM) og en kontinuerlig jumpmodel (CJM). For hver af de 9 aktiver – der dækker aktier, obligationer og råvarer – bruges modellerne til at udpege historiske bullish og bearish perioder. Disse historiske regimer bruges derefter til at træne XGBoost-modeller, en moderne maskinlæringsmetode, der skal forudsige regimet for den næste handelsdag. Analysen bygger på daglige totalafkastdata og evaluerer resultaterne i en periode uden for træning (out of sample) fra 2008 til 2025. Først vurderes, om de forudsagte regimer kan bruges til at forbedre timingstrategier for de enkelte aktiver – altså hvornår man er investeret eller står i kontanter. Derefter indarbejdes regimeregnskaberne i tre typer porteføljer: ligevægtede (EQ), minimum-varians (MinVar) og middel-varians (MV). Disse porteføljer sammenlignes med traditionelle benchmarks, og der tages højde for handelsomkostninger. Begge jumpmodeller kan identificere regimer med forskellige mønstre i afkast og volatilitet, men XGBoost-forudsigelserne forbedrer ikke systematisk resultaterne for de enkelte timingstrategier. Porteføljeanalysen giver også blandede resultater. De tydeligste forbedringer ses i minimum-varians-porteføljerne, hvor JM-XGB og CJM-XGB opnår Sharpe-ratioer på henholdsvis 1,185 og 1,285 mod 0,877 for den traditionelle benchmark, og samtidig har disse strategier markant mindre maksimalt kurstab (maximum drawdown). En analyse af porteføljevægtningen viser, at strategierne i høj grad flyttede midler til kontanter under finanskrisen i 2008 og markedsnedgangen i 2022. Alligevel viser statistiske tests baseret på paired stationary bootstrap, at der ikke er nogen statistisk signifikante forskelle i Sharpe-ratioer på tværs af de otte porteføljer, der sammenlignes. Samlet set tyder resultaterne på, at regimespecifikke prognoser for enkelte aktiver kan forbedre afkast-risiko-forholdet under visse allokeringsregler, men hverken den diskrete eller den kontinuerlige tilgang giver en konsekvent bedre performance end traditionelle porteføljestrategier.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
