The Convenience Yield and the Informational Content of the Oil Futures Price.

AuthorBernard, Jean-Thomas
  1. INTRODUCTION

    Oil prices have exhibited much variability since the first oil well was drilled in Pennsylvania in 1859. Moreover, price fluctuations have been particularly marked since 1973; the level of the series has averaged more than twice its 20$/barrel (in today's dollars) mean over the 1880-1972 period, and as its volatility has increased by almost 50 per cent relative to the pre-1973 era (see the graphs in BP (2013) and the Review of World Energy (2012)). These fluctuations are due to various supply and demand shocks, and are also related to a multitude of geopolitical considerations and the uncertainties that the latter entail. (1)

    With the continuing high volatility in the recent decades, and the expanding use of oil in the global economy, individuals with a vested interest increasingly looked to methods for hedging against future price changes. As a result, the market for oil futures steadily developed, bringing with it an increased trading volume and more liquidity, notably for contracts of maturities of a year or less. In turn, this led policymakers and firms to become more and more reliant on futures prices for their prediction and planning needs. Given such a widespread dependence on the futures market, oil futures would be expected to contain useful information, notably on the future behaviour of spot prices. Yet, recent research has shown that naive no-change forecasts can often outperform forecasts based on futures prices (see, for instance, Alquist-Kilian (2010) and Chinn-Coibion (2013)), which suggests that futures prices do not hold information that improves on random-walk forecasting.

    In this paper we contribute to the above debate by focusing on the extent of information that can be obtained from futures prices when alternative modeling strategies are used to specify the behaviour of oil futures prices. Our analysis is conducted using a forecast-based perspective. More specifically, we study the extent to which futures prices are predictable out-of-sample, based on different available models and in real-time, that is, based on updated parameter estimates for these models ahead of each individual forecast. We show that models which, at each time period, also take into account the difference between the futures price and the spot price (thereby allowing for a time-varying convenience yield), often produce considerably more precise forecasts. In other words, futures prices are found to contain a certain amount of economic and financial information that is useful for forecasting, but more importantly, that more of this information can be reached when both the price level and the distance of the latter from spot price are jointly considered. By extension, we would expect models that rely only on the futures price level (or on the first difference in the level) not to be particularly succesful at outperforming random-walk-based forecasts of spot prices. In this regard, our analysis appears to reconcile, to a certain extent, some of the seemingly contradictory positions and findings in the literature.

    One branch of the literature has relied mainly on structural settings to explain the behaviour of oil prices. Well-known examples of such equilibrium models include dynamic Hotelling-type models for non-renewable resource markets, storage and inventory models for commodity markets, and financial-theory based risk and hedging models for options and futures markets. (2) At the centre of most of these models is the concept of a time-varying convenience yield which can be defined as the flow of goods and services that accrues to the owner of a commodity but not to the owner of a futures contract. (3) It is conjectured that (random) changes in the quantities of held inventories will influence spot prices more than they will affect futures prices, so that the convenience yield will vary over time. Examples of studies that propose models of time-varying convenience yields include, for example, Schwartz (1997), Schwartz-Smith (2000) and Pindyck (1999). In addition, Wu-McCallum (2005) and, to a lesser extent Alquist-Kilian (2010), report that the Hotelling model does relatively well in forecasting the future price of oil.

    Another branch of the literature has used time-series methods to specify oil price dynamics, often relying on statistical testing to determine the model features. In general, futures prices are assumed to follow a random walk, with or without a time-varying variance. To examine whether futures prices are predictors of spot oil prices or whether they perform better than no-change forecasts, statistical tests are applied to the (first-differenced) series within the context of simple linear regression models. Recent findings from this research show that futures prices generally do not outperform no-change forecasts, though results also appear to somewhat depend on the maturity considered for the futures, on the sample period, as well as on the postulated linear relationship (see, for example, Chinn-Coibion (2013) and Alquist-Kilian (2010), as well as the references cited therein). As will be discussed below, implicit in these testing strategies are underlying models where the convenience yield is assumed to be constant.

    It thus appears that the answer to the question of whether futures prices contain useful information on the behaviour of future oil prices is predicated first on the assumed underlying model. In this regard, the empirical literature on oil price modeling reveals that while almost all of the studies acknowledge volatility clusters, heavy tails, and structural breaks, statistical support has been claimed for otherwise fundamentally different types of models. The reasons for such disparity has of course to do with underlying theoretical setups, but also with the various ways data challenges are handled econometrically. The latter include discerning deterministic trends from stochastic ones, disentangling structural change in the fundamentals from inherent fluctuations, using different ways to allow for unexpected discontinuities in price levels or price changes, and accounting differently for non-constancy in the variance of prices and of price changes. Commonly used econometric tools to pin down sample data idiosyncracies include a plethora of unit root and break tests, a battery of generalized autoregressive conditional heteroskedasticity [GARCH] based inference methods, Kalman-filter based time-varying-parameter or neural-network based estimations, or jump-diffusion based procedures. No consensus emerges on what constitutes the best model.

    Importantly, conclusions also depend on whether the convenience yield in the underlying model is assumed to be constant or time-varying. In constant convenience yield models, oil prices are considered to be completely unpredictable, so that a unit root is found to be the best way to represent the behaviour of the mean of the series (the high persistence of oil price has been reported, among others, by Diebold-Kilian (2000)). In studies proposing models with time-varying convenience yield (such as Schwartz (1997), Schwartz-Smith (2000), Pindyck (2001) and Pindyck (1999)) various theoretical and practical reasons are given to suspect that simple unit-root or Geometric Brownian Motion specifications are not appropriate to model natural resources or commodity prices. Instead, these models suggest that prices revert to a mean that is a long-run equilibrium, but that nevertheless this mean may itself change over time due to resource depletion, technological change or product innovations. Reversion to the mean occurs because when prices are higher (or lower) than some equilibrium level, high-cost producers will enter (or exit) the market, which pushes prices downward (or upward). They also conclude for mean reversion by analyzing the relationship between futures prices at different maturities and the spot price, in other words, the time-varying convenience yield. Thus, the convenience yield reflects the information contained in the oil futures prices regarding future values of the spot price and can be exploited. (4)

    In light of the above discussion on the lack of concensus on the best-fitting empirical model, in this paper we follow an agnostic approach to assess whether oil futures contain useful information. Thus, rather than applying various types of tests that, depending on the data sample considered and the idiosyncraticities therein, could have little or no power for some models, or worse, could be oversized and reject spuriously, we prefer to rely on a forecast-based strategy where the exact same criterion is applied to all the models. Importantly, our analysis is also real-timebased; before a new forecast is made, model parameters are first re-estimated and thus updated. We also argue that, in these types of analyses, it is vitally important to consider data of different frequencies, as some models can make better use of the information present in one or the other frequency data.

    Our results show that, with monthly data, forecast performances derived from the different models are largely comparable. However, with weekly frequency data, forecasts based on models that assume time-varying convenience yield are much more accurate compared to forecasts from models that do not make this assumption. This (often substantial) gain in accuracy comes from considering the joint behaviour of spot and futures prices, which shows that deviations at every time period of futures prices from spot prices (reflecting the time-varying convenience yield) are informationally extremely relevant. (5)

    For the same model case, we also document that forecast performances increase with longer date-to-maturity futures. This result suggests that the role of the convenience yield becomes more important with increasing maturity durations, and, by the same token, that longer-maturity oil futures prices have a higher information-to-noise ratio. Finally, we...

To continue reading

Request your trial

VLEX uses login cookies to provide you with a better browsing experience. If you click on 'Accept' or continue browsing this site we consider that you accept our cookie policy. ACCEPT