Guidolin & Timmermann (2007) established that portfolios which ignore regime dependence incur measurable welfare losses for investors. Traditional portfolio construction assumes stable relationships between financial variables. However, the literature has shown that these relationships are time-varying, which leaves conventional frameworks unable to identify attractive assets under prevailing market conditions. This thesis addresses this separation of asset pricing and portfolio optimisation by applying machine learning to 116 firm-level characteristics and eight macroeconomic variables on EUROSTOXX 600 constituent equities. The resulting forward-looking excess return forecasts are incorporated into six portfolio weighting schemes, and performance is evaluated net of transaction costs. To address the regime uncertainty problem, a combination framework equally weighting across all model-scheme pairs is proposed.

The empirical findings show that an ML-based forecasting layer improves risk-adjusted performance relative to non-adaptive full universe benchmarks. However, the value added is reliant on the compatibility between the forecasting layer and the allocation objective, with Risk Parity representing the notable exception. Forecasting quality is shown to be multidimensional. ML models hold a clear advantage over simple benchmarks on level-prediction accuracy, while the cross-sectional ranking accuracy is more competitive, with simple momentum-based forecasting showcasing high ranking ability despite poor level prediction. Moreover, a clear pattern of regime dependence is observed across three distinct market periods examined: stable growth, COVID-19 crash and recovery, and post-COVID-19 uncertainty. In none of these does a single model or weighting scheme consistently produce risk-adjusted outperformance, which confirms the regime uncertainty problem that motivates the combination framework. The combination framework produced consistent risk-adjusted outperformance over the non-adaptive full universe benchmarks across all three regimes, and reduced standard errors relative to individual model-scheme portfolios. It also retained alpha significance after transaction costs when the benchmark forecasting models and naive weighting schemes were excluded from the combination.