Drilling For Value, Appendix B: Commodity Price Forecasts (or, why commodity price forecasts are mostly worthless)

Summary

  • Equity investments into upstream oil and gas companies are largely levered commodity price plays; long-term total returns barely offset the carry costs of taking a long position in oil futures.
  • There are multitudes of ways by which experts seek to forecast future commodities prices; most don’t work.
  • The failure of forecasting should not be surprising if the Efficient Market Hypothesis is even partly correct.
  • Even barring market efficiency, behavioral models provide ample reason for the widespread inaccuracy of forecasts.
  • The idea that commodities prices — including oil — follow a random walk is both overwhelmingly supported by evidence and practical.

Figure 1: Black Gold

Source: Andy Thomas. Black Gold

Evidence overwhelmingly supports the notion that investments into upstream oil and gas producers are basically levered commodity price plays. This, and the fact that commodities producers are price-takers, indicates that petroleum economics are overly levered to commodities prices. It should follow, therefore, that an ability to accurately predict petroleum prices could result in advantageous market timing — i.e. investments in the right petroleum producing assets during the right times in the cycle. As a result of this ostensible potential for riches, prognosticators have devised multitudes of ways to forecast oil prices. Unfortunately, most of these efforts fall short of useful — no known forecasting approach, not even futures strip prices, significantly outperforms the assumption that price evolutions are random walks using out-of-sample data. This failure is not surprising, however, if we are to believe even a watered-down form of the Efficient Market Hypothesis (EMH).

A capitalization weighted index of upstream oil and gas companies exhibits long and short-term price fluctuations that are almost in lock-step with crude oil spot prices. This, and the fact that commodities producers are price-takers, indicates that equity investments into these types of companies are essentially levered commodities plays which have only managed to offset the carrying cost of a long position in crude oil futures. Figure (1), below, handily demonstrates this case.

Figure 2: Upstream Total Return Index vs. WTI Crude Oil PricesSource: Portfolio123

Note to Figure 2: Custom indices are constructed according to a modified capitalization weighted indexing methodology. Upstream Total Return Index includes all U.S. publicly traded companies in the Portfolio123 database which are assigned primary GICS Codes 10102010 (Integrated Oil & Gas) and 10102020 (Oil & Gas Exploration & Production). Dividends explain nearly half of the returns of holding a capitalization weighted index of oil and gas producers. A price return index which excludes dividends would have increased about 58% of the rate of the total return index.

The realization that investing in this space is essentially a casino — unless one can exercise foresight regarding the future direction of commodities prices — leads many to rely on their own or third-party commodity price forecasts. The rationale behind this is pretty straightforward: if one has a probabilistic edge in forecasting commodities prices, this can lead to investments into the right types of assets during the right times of the cycle.

For example, a rational investor armed with foresight would allocate to defensive assets with high margins, long-lived expectancies, and good access to infrastructure preceding downturns, even if those assets command premium valuations. Likewise, our sagastic investor would re-allocate to distressed assets preceding commodity price recoveries, even if those assets are marginally profitable in the present market environment. The net effect of this timing strategy should result in the efficient compounding of capital.

Conventional Econometric Approaches to Forecasting
However, overwhelming evidence indicates that academic and econometric approaches to commodity price forecasts are not very good. Experts are likely to get things wrong, especially the big moves. The most likely explanation for this lack of prescience is that experts are expertly forecasting the past (i.e., telling themselves a story through models which over-fit historical data). Over-fitting is a problem which extends well beyond the realm of commodities prices.

To quote a 2015 post on the fundamentals of the refining business:

With light-sweet oil prices currently hovering around the $45-50/bbl handle, just as recently as last June, BofA analysts had expressed confidence that Brent would continue to trade around $100/bbl1. Last September, EIA expressed with 95% confidence that NYMEX WTI would stay above $80/bbl through January 20152. Last October, Goldman Sachs analysts predicted that WTI would stay above $90 and Brent above $100/bbl for the following three months3.

A recent case study in which asset price researchers falsely believed they had hit pay-dirt is IMF’s June 2014 paper, The future of oil: Geology versus technology. In the paper, researchers presented a strikingly descriptive econometric model which revised previous work in the field with an over-riding assumption: geological constraints impose an upward trend on oil prices, i.e., depletion increases the technical difficulty involved with the extraction of petroleum resources. The model forecasted that oil prices would reach $150 by 2020 due overwhelmingly to geological constraints. While it remains to be seen whether the forecast will work out over the long-run, the authors implemented a complicated demand model and altogether dismissed any possibility of supply shocks which could exert long-term and meaningful effects on price.

In only four short months after the paper was published, the market awoke to the reality that a meaningful supply shock had been underway for the past decade. Since at least 2008, a confluence of market forces had figured out a means to extract economic quantities of petroleum from shale and tight reservoirs, thereby unlocking vast new resource potential. By February 2016, WTI spot prices had bottomed near the $30 handle.

It is clear now — in hindsight — IMF’s premise that supply shocks are irrelevant was false. And it may also be too soon to really know what the geological constraints actually are. Indeed, the North American unconventional resource revolution which came into full bloom over the last decade is beginning to work itself over the United Kingdom and Europe as I write this. Geology appears to be a global phenomenon after all. Drill or Drop? is an independent blog which provides ongoing updates of this phenomenon. In addition, it is widely believed that the South American and the Siberian landmasses contain vastly greater quantities of untapped and unknown petroleum resources.

This is just one of many anecdotes in which economists and academics correctly forecasted the past (i.e., over-fitted the data so that it looked good), but failed to see what was going on right underneath their own noses4.

For brave souls who dare venture into the depths of commodity hell, the ECB’s Economic Bulletin on Forecasting the price of oil provides a solid overview of academic literature on sophisticated approaches, such as risk-adjusted futures, statistical regressions, and structural models.

Figure 3: Monthly WTI Oil Price Expectations based on the Futures and Risk-Adjusted Futures Term Structures

Source: Baumeister and Kilian. Forty Years of Oil Price Fluctuations: Why the Price of Oil May Still Surprise Us. Journal of Economic Perspectives, Vol. 30, No. 1. Winter 2016

Alternative Methods for Forecasting
Given the probable lack of utility of most econometric models for forecasting commodities prices, many practitioners opt to rely on commodities, investment, business, and market cycles analysis.

The idea behind cyclical analysis is fundamentally sound: the drivers of supply and demand are not reducible to linear regressions on a finite number of variables, such as manufacturing capacity utilization — there are vastly more factors at play. Rather than bootstrapping a time-series using explanatory variables, cycles analysis seeks to roughly approximate the market’s present appetite for investment, capacity, risk, and credit in relation to historical levels. Profitable investing does not usually require investors to have a precise time-series which pegs expected value to a number. It does, on the other hand, require investors to accurately conclude that the bias is significantly up, down, or neither. “It is better to be roughly right than precisely wrong” (misattributed to John Maynard Keynes). Oaktree Capital’s Howard Marks might call this method “taking the temperature of the market”.

However, there is little in the way of consensus on how to interpret any given cycle. Some approaches attempt to analyze cycles using price only, some are price-blind, and others utilize some ratio of commodities prices. Aside from variable selection, there are multitudes of confounding factors: that cycles can be of varied lengths and can be embedded within other types cycles invariably leads to the superposition and convolution of factors. But some generalities may exist.

According to a June 2014 paper from the Bank for International Settlements, “business cycles tend to last from one to eight years, and financial cycles around 15 to 20 years.” Even though financial cycles are increasingly globalized, business cycles can still be intensely local in nature. Furthermore, there may exist a preponderance of commodities super-cycles which, for oil, “have historically lasted an average of 27 years, with a minimum period of 21 years and a maximum of 32 years”.

The lengthy nature of cycles combined with the lack of data means that many perceived historical patterns will be spurious. Clinical psychologists have a name for this: apophenia is the tendency for the human brain to make out patterns where no pattern exists — in extreme cases, it can lead to diagnosed mental illness. Moreover, in an infinitely large data-set, there will be an infinite number of convincing but spurious correlations. And as we all already know: correlation does not imply causation.

There is also the problem of articulation within a broader analytical framework. Multi-variate models for market cycles ultimately fall prey to the same fallacies as complex econometric models as soon as factors become ranked, weighted, parameterized, calibrated, and recalibrated. Unlike within the fields of physical science in which constants are “constant”, markets are functions of non-stationary and non-linear processes. Even if one could identify a correct model for the underpinnings of market behavior, there is the insurmountable problem associated with the accurate measurement of non-stationary, non-independent, and non-linear variables.

In spite of the drawbacks to cyclical analysis, any modicum of study on market cycles reveals a common central theme: commodities prices are cyclical. ‘Nuff said.

The Problem with Efficient Markets
The failure of conventional models should not come as a huge surprise to those that subscribe to even the weakest form of the EMH. Even if markets used to be relatively more inefficient, the proliferation of capital now means that the financial industry pays the best and attracts the best. This has changed the game.

For example, over any sufficiently long time-frame, the overwhelming majority of active money managers under-perform even their self-selected benchmarks. If professionals who went to the best schools and had the best professors cannot do better than average, what does that portend for lay-folk, such as I claim to be? Moreover, if the marketplace reflects the capital-weighted consensus of all known information at any given point in time, there is still great fundamental uncertainty regarding geology and future technological advances with which to contend.

The Problem with Inefficient Markets
Even if one does not presume that markets are efficient, research in behavioral economics provides ample rationale as to why experts are likely to get things wrong. In spite of their advanced degrees and high IQs, experts may be particularly susceptible to certain kinds of cognitive dissonance.

Experts tend to be particularly predisposed to cognitive biases which lead them to overstate confidence in their own judgments. For example, experts are likely to become caught up in intelligence traps (i.e., overconfidence biases) whereby false opinions are supported through a narrative which is carefully constructed to match selected observations. In the mildest form, overconfidence biases lead to misjudgments in  certainty, i.e., “I am an expert therefore I am more likely to be right”.

Well informed persons are furthermore susceptible to the bandwagon effect. Individuals within a group often tends to accept that group’s consensus due to the safety in numbers principle (i.e., how can everyone be wrong?). For those who follow a given market regularly, it is very difficult not to become infected by the latest vogue.

Bandwagon effects tend to distort and exacerbate commodities and investment cycles and, in extreme cases, result in speculative bubbles. For example, in bull markets, increasing prices increases buying interest for no other reason than that enough people believing in a fallacy can make it true (i.e., “the trend is your friend”). In fact, the presence of momentum in most financial markets is probably the most problematic phenomenon to those subscribe to the notion of market efficiency because it suggest that there is a free lunch to be had — unrelated to any kind of risk premia — by simply buying winners and selling losers.

Attempts to rationalize irrational behavior is clearly an exercise in futility5.

Therefore, it might follow that in order to be a good forecaster, one must be willing to hold unpopular views. But even if one could accurately forecast economic conditions according to some model, that model would invariably fail to predict random events which result in events such as unanticipated disruptions. Event-driven risks foil the efforts of even the best prognosticators.

The Practitioner’s Dilemma
Even good economic forecasts may not be very useful to investors of upstream companies or their managements. From a capital allocation perspective, economic forecasts may be meaningful to select the most advantageous sectors or products to invest in. However, if constrained to investments in assets or securities within the upstream petroleum sector, macroeconomic factors are mostly extraneous since neither managements nor investors exert meaningful influence over them. Given the difficulty of forecasting prices, and probable lack of incentive for doing so, it could said that any such effort is a waste of time and energy.

Against this back-drop of all-encompassing futility, decisions which are affected by uncontrollable and unforeseeable forces must still be made. Companies, regulators, and individuals must have some expectancy of the future in order to press forward.

For example, it is standard practice for integrated petroleum companies to evaluate capital budgeting and investment decisions using a “most conservative” case for oil prices. These planning assumptions are likely indicative of a given company’s resilience to downturns, and may also be inversely indicative of gearing to commodity price upside.

Table 1: Commodity Price Projections

Company Price Projection for Investment Decisions
BP $50
Chevron $52
ExxonMobil $40
Royal Dutch Shell $45*
Total $40

Note to table 1: Applies only to Shell’s deepwater production assets; data on other production sources not given.

Source: Tyler Crowe. 5 Point Checklist for Investing in Big Oil. Motley Fool. 26 Jun 2016. Original content accessible: https://www.fool.com/investing/2016/06/26/5-point-checklist-for-investing-in-big-oil.aspx

It is generally a best practice to evaluate expected returns on capital using a range of commodities prices. This, however, seems to vex many upstream industry professionals who tend to use a single price deck in evaluating well and field-level economics.

Random Walks as a Practical Framework
The intuition to rely on non-deterministic estimation methods is strongly supported by evidence that oil prices follow a random walk. A random walk — to you and me — does not necessarily mean that the underlying dynamics of the supply, demand, and price equilibria are themselves random. It simply means that the end result is virtually indistinguishable from a random process.6

The most common type of continuous random walk used in asset price models is Geometric Brownian Motion; slightly more realistic models, such as exponential Ornstein-Uhlenbeck processes, incorporate mean-reverting properties7. Even more sophisticated model incorporate jumps, which usually assume the distribution of a Poisson process.

In the presence of arbitrage relationships, a risk-neutral distribution is almost always the correct one. Yet there are multiple types of processes which can result in no arbitrage. Moreover, almost every type of continuous random can be calibrated to sufficiently fit observed price distributions. What is not important: selecting the correct process. What is important: selecting a process which reflects one’s beliefs about the underlying drivers of returns as well as sources of uncertainty. There is also the problem of calibration: more sophisticated models are more difficult to parameterize and are subject to greatly more over-fitting risk.

If oil prices were not a random walk, prices from the futures term structure (i.e., futures strip prices) should approximate future prices better than the spot price adjusted for inflation and volatility (i.e., quadratic variation)8. However, term structures do not provide significantly more accurate forecasts than a random walk.

No improvement over the spot price forecast is in fact the expected result if term structures simply reflect expected carry (i.e.,storage) costs — according to Fama and French (FF) (1987), the carry theory of futures pricing is not controversial. While FF still acknowledge the possibility of the existence of risk premia (i.e., expectations of future prices embedded into market prices but independent of carry costs), their presence is controversial and unproven9.

Moreover, no improvement over the spot price forecast is also expected by the no-arbitrage argument of the EMH. According to this argument, arbitrage is the casual mechanism for market efficiency: wrong prices result in arbitrage; arbitrage results in trading; trading eliminates the arbitrage. The converse also holds: the absence of arbitrage implies that prices are not wrong (but does not necessarily imply that they are right).

It is also notable that most types of random walks assume finite variation — i.e., they cannot easily anticipate price regime changes. But, as we have discussed, paradigm shifts which are driven by real world events and technological breakthroughs do happen. This perhaps explains why so many experts are wrong: experts (should) know that it is nearly impossible to improve upon naive extrapolation so, when there is a level change, it always comes as a surprise.

Or, in more eloquent words:

It isn’t surprising that experts aren’t good at predicting prices. Global oil markets are a function of countless variables — geopolitics, economics, technology, geology — each with its own inherent uncertainty. And even if you get those estimates right, you never know when a war in the Middle East or an oil boom in North Dakota will suddenly turn the whole formula on its head.

– Ben Casselman. The Conventional Wisdom On Oil Is Always Wrong. FiveThirtyEight. 18 December 2015

Presented with a random walk, it is a good first approximation for rational investors to assume that “the most likely price in the future is the price today”. Dealing with the range of possible outcomes is only slightly more difficult. But again, it is also a good approximation to estimate financial outcomes given a conservative floor and ceiling on the underlying revenue and cost drivers. More sophisticated methods are distinguished mostly by their consistency with theory, but may not always be more practical10. Many practitioners in fact rely on heuristic models and approximations11.

Any further effort spent forecasting is, for all the reasons cited above, likely to suffer from its own sort of diminishing returns.

Footnotes   [ + ]

1. Crude oil benchmark backwardation to face pressure moving forward: BofA ML
2. EIA, Short-term Energy Outlook, September 2014
3. U.S. Oil Output May Slow If WTI Drops Below $90: Goldman
4. The IMF’s prediction regarding the impending end of the petroleum-age appears to be convenient. This positions is consistent with the IMF’s not-so-implicit agenda of imposing a global tax regime on carbon. The justification seems altruistic enough: anthropogenic climate change is the greatest impediment to a future of sustainable prosperity. However, it is perhaps too convenient that a global carbon tax is being promoted as the panacea for all the woes of society. Will it really solve any core problems such as growing irrelevance of labor capital, over-population, or unworkable behemoths of bureaucracy? Is the premise of a global tax is at odds with classical economic ideals? Moreover, is the almost singular emphasis on carbon a red herring which belies more pressing problems regarding environmental damage caused by real pollution (e.g., ammonia and phosphorous, sulfuric and nitric acids, formaldehyde, ozone, etc…)?
5. It is perhaps indicative of underlying fragility that our present financial system is built on models which assume humans behave like rational agents (i.e., “econs”) according to some idealized utility function.
6. Random processes in financial markets have been studied since Louis Bachelier published Théorie de la spéculation in 1900. Bachelier’s framework exposed the properties of a type of random walk now known as Brownian Motion.
7. Pindyck, Robert S. The Long-Run Evolution of Energy Prices. The Energy Journal, vol. 20, no. 2, 1999, pp. 1–27. JSTOR,
8. Quantitative types refer to spot price forecasts as semi-martingales. A semi-martingale is the expected value of a finite variance process. In the common case of Geometric Brownian Motion (GBM), the unconditional expected value is a continuous function of the current price multiplied by an inflator for the force of interest. Marginal improvements over GBM might be had by adding terms for jump stochasticity and mean reversion, but these terms also partly offset one another. Most types of semi-martingales used in financial modeling assume that past price changes may contain only probabilistic information regarding future price changes.
9. Hamilton and Wu (2014) propose one method for extracting the risk premium for futures price data. However, sophisticated methods such as this which seek to decode this information are prone to over-fitting risk.
10. Since at least the 1960s, culminating with the Black-Scholes’ derivation of the continuous risk neutral measure in 1973, financial math has provide reduced form means of dealing with time value of money under uncertainty (i.e., conditional expectations of random walks). Reference: Fischer Black, Myron Scholes. The Pricing of Options and Corporate Liabilities. The Journal of Political Econony, Volume 81 Issue 3 (May-Jun 1973), 637-654
11. Espen Gaarder Haug, and Nassim Nicholas Taleb. Option Traders Use (very) Sophisticated Heuristics, Never the Black–Scholes–Merton Formula. Journal of Economic Behavior and Organization, Vol. 77, No. 2, 2011. 16 Nov 2012