Crude oil arbitrage and momentum trading strategies with targeting volatility by using Large Language Models

This paper is devoted to the execution of crude oil arbitrage and momentum strategies with volatility targeting by using Large Language Models (LLM), comparing efficiency and profitability of LLaMA2 and GPT3.5. The objective of this study is to estimate the impacts of oil price volatility on oil companies' strategic investments from a regional level. Therefore, oil price volatility should be an independent variable in the model equation. This method is consistent with earlier studies that investigate the relationship between uncertainty and investment in Tobin's q theory. Oil price uncertainty can arise from a number of different sources including global oil demand and supply conditions, the actions of institutional actors (OPEC), geopolitical issues (approximately 50 per cent of world proven oil reserves are located in just four countries in the Middle East), and speculation in oil future markets. The complicated elements contribute the oil price uncertainty to be high and hard to predict. Therefore, although high oil prices normally imply high profits for oil exploration and production companies, the uncertainty of price could still significantly impact their strategic investment decisions. Previous studies diversify their results in the shape of the relationship between oil price uncertainty and investment. For instance, studies such as Mohn & Misund (2009) and Elder & Serietis (2010) find a linear relationship between oil price uncertainty and investment; while Henriques & Sadorsky (2011) found a Ucurve relationship. Therefore, in this study, the variable of oil price volatility square will be introduced as an optional variable. If the specification of model with the variable shows a better result than the one without it, then the shape of the relationship would be a curve line. Volatility targeting is a portfolio management tool aimed at managing portfolio risk by targeting a particular volatility level and adjusting the positioning of the portfolio in an attempt to stay close to the volatility target. Although it is possible to implement volatility targeting with any asset, portfolios of derivatives provide a cost-effective implementation as trading is often done on a daily basis. Following the research of Moreira and Muir (2017) the idea is very straightforward; one needs to simply scale asset weights in the portfolio by the ratio of the target volatility and the expectation of future volatility given the current asset mix. Obviously, the value of the targeting is greatly impacted by the quality of the volatility expectation model, but since these estimates are over short horizons, the fact that volatility tends to cluster makes this an "easier" effort. In equation form, the return of the volatility managed portfolio, is the volatility scaled return of the unmanaged portfolio which is simply the sum product of the asset weights and their returns. Crude oil is a pivotal global commodity whose price fluctuations permeate the broader economy through channels such as transportation costs, energy-weighted inflation indices, and corporate procurement and inventory strategies. Accurate forecasting of oil prices, even for single-step daily predictions, holds significant practical value. Marginal but consistent accuracy gains around price turning points can substantially mitigate risks associated with supply disruptions and excessive hedging costs. However, the inherent characteristics of oil markets-such as fat-tailed residuals, volatility clustering, asymmetric responses to supply and demand shocks, and frequent structural breaks-pose considerable challenges to traditional linear models. While classical time series approaches like ARIMA and exponential smoothing capture low-frequency trends, they often fall short in modeling the complex, nonlinear interactions among inventory levels, convenience yields, market expectations, and geopolitical risks. Although regime-switching models introduce nonlinearity, their reliance on strong assumptions regarding the number of states and transition mechanisms limits their applicability during unconventional events. From a methodological viewpoint, I improved on existing approaches to pairs trading by assessing the cointegration among the three futures prices at the same time and modeling the resulting cointegration spread by a mean-reverting process with regime-switching modulated by a hidden Markov chain. This generalizes the approach of Tenyakov and Mamon, 2017 and Elliott and Bradrania, 2018 and enables us to