When GARCH is not enough: Volatility dynamics and extreme tail risk in coffee commodity markets

Standard risk models struggle with commodity markets. This paper asks whether that limitation holds for a listed agricultural commodity ETC, and whether semi-parametric methods can overcome it. Using nearly twenty years of daily closing prices from the WisdomTree Coffee ETC (COFF.MI; ISIN: JE00BN7KB557), listed on Borsa Italiana from September 2006 to March 2026, we characterise the statistical and risk properties of a financialised agricultural commodity instrument. The price series conforms to the stylised facts of financial time series — non-stationary prices, stationary returns, pronounced volatility clustering, and fat-tailed return distributions. Among six GARCH-class specifications estimated and compared, the GJR-GARCH model with Student- t innovations is selected as the preferred specification, capturing both size-asymmetric volatility responses and extreme tail behaviour that places the return distribution at the boundary of finite variance and infinite kurtosis. The central finding concerns risk measurement. Rolling-window out-of-sample backtesting shows that all three parametric specifications tested — including the preferred asymmetric model — fail to provide adequate 99% Value-at-Risk coverage, with violation rates of 1.66% to 5.07% against the nominal 1% target. A semi-parametric Extreme Value Theory approach, which models the tail of the empirical loss distribution directly without parametric assumptions, is the only method to achieve adequate coverage (violation rate 1.44%, not rejected at the 1% level), while also providing the first empirical estimate of Expected Shortfall for this asset class (mean 7.97%). Supplementary analysis finds no intra-annual seasonality — consistent with the globally distributed production calendar of coffee — and identifies a dominant medium-term price cycle of approximately 40 months, coherent with the biological production lags of perennial crops. Together, the results demonstrate that the financialised ETC retains the empirical characteristics of the underlying commodity market, and that measuring its tail risk requires methods robust to the structural instability that defeats standard GARCH-based approaches in real-world forecasting settings