This thesis investigates commodity risk premia in the context of modern portfolio management. It focuses on portfolio allocation methods to better capture the risk premia, the strategic role of commodity risk premia in multi- asset portfolios, and whether sophisticated modelling techniques, including machine learning methods, can be used to exploit information from the state of the economy and financial markets to capture the commodity risk premia. In the first empirical study presented in Chapter 2, we examine whether more sophisticated weighting schemes lead to stronger performance of commodity investment portfolios due to an enhanced capacity to capture risk premia. We go beyond the naive equally weighted portfolio that dominates the commodity literature, and consider sophisticated portfolio allocations (i.e., utility optimization, mean variance optimization, risk minimization, and risk timing) to improve the performance characteristics of the investment portfolios. Portfolios are long-short and are constructed using momentum or term structure signals. Three variations of utility function optimization (i.e., power utility, power utility with disappointment aversion, and negative exponential utility) and several variations of risk timing methods (i.e., volatility timing, reward-to-risk timing, beta timing, Value-at-Risk timing, and conditional Value-at-Risk timing) are implemented. We show that risk timing and risk minimization weighting schemes lead to desirable portfolio characteristics such as greater risk-adjusted returns for most of the investigated portfolios. In addition to momentum and term structure signals, strong results persist when alternative methods are used based on hedging pressure or basis momentum in the formation of long-short portfolios. We find the results to be robust to transaction costs, liquidity risk and the choice of model parameters and confirm they are not an artefact of data mining. The second empirical study is presented in Chapter 3. In this study, we assess the benefits of a strategic allocation to commodities in multi-asset portfolios. Traditional portfolios of equities and bonds are constructed using various exposures and allocation methods. The traditional portfolios are then augmented with commodity exposures, and the performance of these combined portfolios are measured and compared against their traditional counterparts that exclude commodities. A range of commodity allocations are considered which include naive and simple exposures, to act as benchmarks, and more sophisticated long and long-short portfolios based on integrating various investment styles. More specifically, an integrated signal consisting of momentum, value, term structure, skewness, and speculative pressure stimuli is responsible for the sophisticated commodity allocations. With respect to equities and bonds, various allocations are considered including traditional and widely-used indices, style-integrated futures portfolios, and equity style portfolios such as size, value, and trend-following portfolios. Different portfolio weighting schemes (i.e., naive equally weighted, weights based on utility maximization, and risk timing) are implemented to assess the importance of allocation methods. The results show that investment portfolios that include style-integrated commodity exposures exhibit improved out-of- sample risk-adjusted returns and drawdown measures that are statistically significant compared to traditional portfolios. The benefits are present for both long-only and long-short allocations, irrespective of the type of equity or fixed income allocations. Furthermore, naive commodity allocations fail to add value to traditional portfolios. Chapter 4 presents the third and final empirical study, focusing on capturing commodity risk premia using a large dataset of macroeconomic and financial variables. Specifically, we construct linear and nonlinear predictive models to study the linkage between 128 macroeconomic and financial predictors representing the state of the economy, an unobservable variable, and subsequent commodity futures returns. Linear models consist of predictive regression models employing two shrinkage methods to reduce the dimensionality of the large predictor set. Shrinkage is done either naively by producing one factor from the standardized average of the predictors or by selecting the first few principal components of the predictors in lieu of the full cross section. Nonlinear techniques consist of machine learning methods, centred on two alternative approaches. One alternative utilizes a deep neural network, while the other combines the deep neural network with a recurrent neural network, namely, the long short-term memory network. Based on positive results from all four methods, we conclude that the state of the economy carries information about commodity risk premia which can be realized via constructing profitable portfolios. We also show that the performance is strongest for the combination of feedforward and recurrent networks and that the risk premium is unrelated to, and exceeds, those earned on previously- published characteristic-sorted portfolios. We show our results are robust to liquidity risk, transaction costs, and positive performance persists across various market environments.