Corporate Risk Management: The Hedging Footprint. PETER MACKAY and SARA B. MOELLER ABSTRACT

Transcription

1 Corporate Risk Management: The Hedging Footprint PETER MACKAY and SARA B. MOELLER ABSTRACT What risks do firms hedge? How much do they hedge? How far ahead do they hedge? What determines corporate hedging policy? Should firms hedge at all? Can corporate risk management create value? As straightforward and important as they might appear, these questions are still largely unresolved. One difficulty in answering them is lack of data: Largesample studies rely on coarse measures (does a firm hedge or not?) that offer few insights and studies with precise measures rely on small-sample, proprietary data that do not generalize. We propose an alternative approach that extracts corporate hedging policy from publicly-available data. The key insight is that the way corporate risk-management activity is recorded (cash-flow hedge accounting) leaves a hedging footprint that we can uncover by regressing a firm s sales (or costs) against past futures prices and recent spot prices. We apply our approach to a sample of 34 oil refiners and return to answer some of the above questions. Preliminary, please do not cite or circulate. Mackay is from the School of Business Management, Hong Kong University of Science and Technology and Moeller is from the Babcock Graduate School of Management, University of Pittsburgh. We thank Ilona Babenko, Jonathan Batten, Magnus Dalqvist, Sudipto Dasgupta, Michael Gordy, Vidhan Goyal, Delroy Hunter, Mike Lemmon, Laura Liu, Lars Oxelheim, Mark Seasholes, Per Strömberg, Sergey Tsyplakov, Chu Zhang, and seminar participants at the Hanken School of Economics and Helsinki School of Economics, the Hong Kong University of Science and Technology, the Swedish Institute for Financial Research, the University of South Florida, the Asia-Pacific Association of Derivatives (Korea), the NUS Risk Management Conference (Singapore), and the International Risk Management Conference (Venice), for useful comments and discussions. 1

2 Corporate Risk Management: The Hedging Footprint ABSTRACT What risks do firms hedge? How much do they hedge? How far ahead do they hedge? What determines corporate hedging policy? Should firms hedge at all? Can corporate risk management create value? As straightforward and important as they might appear, these questions are still largely unresolved. One difficulty in answering them is lack of data: Largesample studies rely on coarse measures (does a firm hedge or not?) that offer few insights and studies with precise measures rely on small-sample, proprietary data that do not generalize. We propose an alternative approach that extracts corporate hedging policy from publicly-available data. The key insight is that the way corporate risk-management activity is recorded (cash-flow hedge accounting) leaves a hedging footprint that we can uncover by regressing a firm s sales (or costs) against past futures prices and recent spot prices. We apply our approach to a sample of 34 oil refiners and return to answer some of the above questions.

3 A large literature studies the motives for risk management and the extent of hedging activity. 1 Most of this literature deals with the intensity of hedging; very little has been said on how far ahead firms should hedge and what contracts to use among the several available delivery dates. Yet, managers must make specific decisions when it comes to implementing hedging strategies. 3 Compounding the lack of theoretical guidance, we know nearly nothing about the determinants and value implications of the maturity structure of corporate hedging in practice. Data limitations are largely to blame since even a careful reading of financial-statement footnotes rarely yields more than a coarse measure of hedging activity, let alone details such as contract maturities. This study proposes a way to back out the intensity and maturity of corporate hedging from limited data, which allows us to investigate these important dimensions of corporate risk management. Building on FASB accounting rules for derivatives, we propose a simple method to extract corporate hedging activity from readily-available data. Briefly, gains or losses on derivatives used to hedge a firm s sales or costs are deferred until the related transaction is actually reported. Corporate risk-management activity recorded in this way (cash-flow hedge accounting) leaves a footprint that maps a given period s sales and costs to the time path of its hedgeable risk factors. We can therefore uncover a firm s hedging policy by regressing its sales and costs on the lagged futures prices and recent spot prices the firm faced on its product and factor markets. Following MacKay and Moeller (007), we apply our approach to a sample of 34 oil refiners by regressing quarterly sales and costs on NYMEX energy prices from March 1985 to June 004. We then study how well the estimated intensity, maturity, and value of corporate risk management agree with footnote-based estimates and relate to theoretical predictions and futures-market conditions. 1 3 A seminal paper by Smith and Stulz (1985) lays the ground work for subsequent theory and related empirics. Smith and Stulz show that by stabilizing cash flows, corporate hedging can add value when firms face market imperfections that result in nonlinear payoffs (e.g., progressive taxation, bankruptcy costs). Their central idea that nonlinearities justify hedging has since been applied to other financial factors such as costly external finance (Froot, Scharfstein, and Stein, 1993) and information asymmetry (DeMarzo and Duffie, 1991). MacKay and Moeller (007) apply Smith and Stulz s model to estimate the value of corporate risk management related to factors affecting the real side of the firm. See MacKay and Moeller (007) for a review of the empirical literature. One recent exception is Fehle and Tsyplakov (005), who model both the optimal hedge intensity and maturity. Neuberger (1999) shows what mix of contract maturities minimizes rollover risk when the desired hedge horizon exceeds the longest-dated futures contract available. The notorious Metallgesellschaft debacle is a case in point (see Mello and Parsons, 1995). 1

4 Before turning to our results, let us elaborate a little on our approach through an example. Suppose a firm hedges half its price exposure one year ahead of delivery. Its reported sales (or costs) will therefore reflect two prices: the recent spot price and the futures price it contracted a year earlier. Our approach lets the data speak by allowing the estimation to endogenously assign weights to the spot price and the six lagged futures prices we consider, i.e., 3, 6, 9, 1, 18, and 4 months. 4 Thus, for the firm just described, our approach would assign a weight of 50% to the spot price and 50% to the lagged 1-month futures price; all other lagged futures prices would receive weights of zero. 5 The resulting decomposition tells us about the level of hedging activity (the sum of the weights assigned to the lagged futures prices) as well as the maturity structure of corporate hedging (the distribution of the weights assigned across the lagged futures prices). Our results can be summarized as follows. First, the average sample firm hedges about a quarter of its exposure. Specifically, our sales (costs) regression assigns 4% (%) of the weight on the combined set of lagged futures prices and the remaining 76% (78%) on spot prices. This level of hedging activity is higher (but not significantly) than the level obtained if we only include the lagged three-month futures prices (as in MacKay and Moeller, 007), which shows that including longer-dated contracts improves the specification and helps to correct a downward omitted-variable bias in the estimated level of hedging activity. Thus, we argue that the longer-dated contracts belong in the model on both economic and econometric grounds. Second, we find that the average sample firm hedges well beyond the nearest quarter, with a significantly shorter hedge maturity for sales (about 7 months) than costs (about 11 months). However, these results vary considerably when estimated at the firm level, both in comparison to 4 5 Just to be clear: The lagged three-month futures price means the average price of a three-month maturity contract observed in the quarter prior to the current quarter; the lagged six-month futures price means the average price of a six-month maturity contract observed two quarters prior to the current quarter, etc. The longest maturity we consider is 4 months, so the lagged 4-month futures price means the average price of a 4-month maturity contract eight quarters prior to the current quarter. In reality firms pursue much more complex, time-varying hedge strategies, which will introduce estimation error. Our approach also assumes that the use of futures contracts qualifies for so-called hedge accounting and that a firm s entire risk management program reduces to its use of futures contracts. We return to these questions later.

5 the sample average estimates and across firms. For instance, the median firm-level estimates of hedge intensity are 46% (49%) for sales (costs) and (14.7) months for hedge maturity. Third, our results are somewhat sensitive to what specification is used, in particular, which of the six contract maturities are included in the estimation. For instance, in a regression that includes the lagged futures prices for 3- and 6-month maturities, the lagged 9-month futures prices enter significantly, both economically and statistically (weight of 1% and confidence levels of 5% or higher). However, this weight is shifted away from the 9-month maturity and onto the 1-month maturity once the lagged 1-month futures prices are added to the regression. At this point the model stabilizes, even as we add the 18-month and 4-month maturities. Fourth, our analysis of which contract maturities belong in the model which of the six lagged futures prices are assigned significant positive weights reveals that the average sample firm does not hedge every point along the two-year horizon we consider. For instance, in a regression that includes all six contract maturities, only three enter significantly, namely, three months (with statistically significant weights of 11% for sales and 10% for costs), one year (only sales has a significant weight of 10%), and two years (only costs has a significant weight of 6%). In other words, the average firm exhibits a non-contiguous, asymmetric maturity structure that skips the 6-, 9-, and 18-month maturities and hedges sales up to one year ahead and costs up to two years ahead. Dropping these non-significant contract maturities from the regression places fewer demands on the data without reducing goodness of fit. This model, which only uses spot prices and the 3-, 1-, and 4-month lagged futures prices, is the one we use at the individual firm-level to estimate the hedge intensity and hedge maturity for our 34 oil refiners. While our approach might seem reasonable, it certainly has its limitations and is subject to estimation error, which may lower its usefulness. We therefore use a variety of strategies to validate the approach. First, given the small size of the sample, we search each firm s financialstatement footnotes for clues on its stated hedging policy. We allow for three levels of hedge intensity: Usually hedges, sometimes hedges, or never hedges. For hedge maturity, we note the 3

6 longest contract maturity the firm states that it normally uses. Based on bootstrapped clusterederror estimation, we find that the correlation between our estimated measures of hedge intensity and hedge maturity and those gleaned from footnotes are respectively 18% and 11% for our sales-based measures and 15% and 17% for our cost-based measures. Second, we are able to replicate key results reported in Fehle and Tsyplakov (005), where hedge intensity and hedge maturity are more precisely constructed from detailed data for firms in the gold industry. Finally, we propose a number of determinants of hedging policy (discussed below) and obtain regression results that conform to our hypotheses, which again helps to validate the approach. So, what are the determinants of hedging policy? We first test empirical predictions from Fehle and Tsyplakov (005), namely, that both hedge intensity and maturity should increase then decrease with the probability of financial distress. We regress our hedging-policy measures on Altman s Z-score, the squared value of Altman s Z-score, and control variables. Then, for robustness, we use financial leverage instead of Altman s Z-score as a proxy for the probability of financial distress (Fehle and Tsyplakov only use financial leverage). Results are mixed across proxies for the probability of financial distress, between hedge intensity and hedge maturity, and for sales-based versus cost-based hedging-policy measures. Contrary to Fehle and Tsyplakov, regardless of proxy (Altman s Z-score or financial leverage), we find an inverse relation between hedge intensity and the probability of financial distress. But consistent with their results, we find that hedge intensity decreases as the marginal probability of financial distress increases. Results for hedge maturity are even more nuanced: As in Fehle and Tsyplakov (005), regardless of proxy (Altman s Z-score or financial leverage), our sales-based measure of hedge maturity first rises then falls with the probability of financial distress. But for our cost-based measure, regardless of proxy, hedge maturity is inversely related to both the probability of financial distress and its squared value. Thus, we only partly corroborate Fehle and Tsyplakov. Maturity matching is a common prescription and practice in corporate finance, whereby firms seek to match the maturities of their assets and liabilities both on and off the balance sheet. 4

7 We therefore hypothesize a positive relation between hedge maturity and the maturity of a firm s assets as well as the maturity of its financial structure. In other words, we expect that firms with a greater fraction of short-term assets and short-term debt will use hedges of shorter maturity. Our results support these predictions: hedge maturity falls significantly with the ratio of net working capital to total assets and with the ratio of short-term debt to total assets. We find mixed results for risk: Hedge maturity falls with cash-flow volatility but rises with diversification. We also investigate the determinants of the level of hedging activity itself. For instance, we find that large, profitable, diversified, low cash-flow risk firms hedge less, which supports the idea that firms with better access to external capital markets have less need to hedge. We also explore the relation between corporate risk management and organizational form. We find that vertically-integrated firms use significantly shorter hedge maturities, consistent with the notion that upstream and downstream business segments serve to internalize price swings and effectively shorten the price exposure maturity of integrated oil refiners. Second, we find that diversified firms hedge about one-third as much as less diversified firms, consistent with our finding that diversified firms also derive significantly less value from risk management. Finally, we find that corporate risk management depends on futures-market conditions. For instance, hedge intensity more than halves when the reference price (the difference between output price and input price, also known as the crack spread ) rises above the sample median. Even more striking, hedge intensity is about four times higher when price volatility is low than when it is high. These results suggest that oil refiners have difficulty locking in high spreads and finding counterparties when price volatility is high. Hedging intensifies when backwardation and term-structure risk rise, which suggests that refiners attempt to exploit inconsistencies along the term structure. Hedge maturity is much shorter when open interest is skewed toward the longdated contracts, which further corroborates the idea that refiners try to time the futures market. This paper makes three contributions. First, we propose a straightforward way to backward engineer corporate hedging activity. Our approach uses standard COMPUSTAT data rather than 5

8 the usual, problematic data sources, and thus represents an important methodological innovation. Second, we provide some of the first descriptive evidence on corporate maturity structure. While our results are specific to oil refining, the approach can be applied to other industries. Finally, we examine the determinants and value implications of hedging policy. We corroborate key results in Fehle and Tsyplakov (005), which supports their predictions and validates our approach. The rest of this paper is structured as follows. Section I motivates the study and develops the approach. Section II describes the data. Section III presents summary statistics on energy prices, firm characteristics, and derivatives usage. Section IV presents industry- and firm-level estimates of hedge maturity and hedge intensity. Section V examines the determinants and value implications of hedging policy. Section VI concludes. I. Motivation and Methodology One of the challenges facing empiricists is the lack of detailed data pertaining to corporate hedging activity. Researchers generally have resorted to one of three approaches: 1) Estimate an extended market model that includes the return on the risk factor of interest (e.g., Flannery and James, 1984, for interest rates, Jorion, 1990, for foreign exchange rates, Strong, 1991, for oil prices, and Tufano, 1998, for gold prices). Unfortunately, although this adhoc approach might interest diversified investors to the extent that such risk factors are priced, it subsumes the information relevant to corporate risk managers because stock returns are net of corporate hedging activity. ) Collect detailed data for a small set of firms, usually for a single industry such as gold mining, through surveys or proprietary data (e.g., Tufano, 1996, 1998, Haushalter, 000, Brown, 001, Haushalter, Heron, and Lie, 00, Adam and Fernando, 006, Carter, Rogers, and Simkins, 006, Brown, Crabb, and Haushalter, 006, Jin and Jorion, 006). Unfortunately, this approach is very labor-intensive and it is not clear whether the results extend to other firms or industries. This approach may be free of estimation error but is still prone to measurement error. 6

9 3) Search through the financial-statement footnotes of a large sample of non-financial firms, a procedure that often yields no more than a hedge versus no hedge dummy variable (e.g., Nance, Smith, and Smithson, 1993, Mian, 1996, Geczy, Minton, and Schrand, 1997, Guay, 1999, Allayannis and Weston, 001, and Hentschel and Kothari, 001). However, firms are now required to report greater detail on their derivatives positions, which some authors have exploited to produce continuous measures of hedging (e.g., Graham and Rogers, 00, Guay and Kothari, 003, and Bartram, Brown, and Fehle, 006). Unfortunately, even with more stringent disclosure rules in recent years, our own experience shows that actual disclosures fall short of requirements: Firms may only discuss their use of derivatives qualitatively rather than quantitatively; reported numbers are often notional amounts rather than position values; and firms typically do not distinguish between derivatives used to hedge rather than speculate. A. Description of the proposed approach Our proposed approach is simple to implement and uses nothing but standard financialstatement data, such as those found in COMPUSTAT, and market-traded derivatives prices, such as those found on Thomson Reuters DATASTREAM. The main insight is that the accounting treatment of derivatives used to hedge firm cash flows (known as a cash-flow hedge) leaves a detectable footprint. We first explain how cash-flow hedge accounting works then how it leads to a straightforward way to detect the corporate hedging footprint using readily-available data. Our approach rests on two features of cash-flow hedge accounting. 6 First, gains or losses on derivatives positions that qualify as cash-flow hedges are reported in sales or costs rather than under other income, where non-qualified derivatives gains or losses are reported. 7 Second, qualifying gains or losses only appear on the income statement in the period when the underlying 6 7 The development we provide here closely follows Kieso, Weygandt, and Warfield, 008 (pp ), a widelyused financial accounting textbook. The related FASB directive is 133 (issued in 001). Derivatives positions used to hedge balance-sheet accounts, rather than sales or costs, are called fair-value hedges (e.g., interest rates swaps used to fix borrowing costs). Fair-value hedges and all other uses of derivatives that do not qualify as cash-flow hedges follow normal GAAP (e.g., hedges meant to protect a competitive position, time the market, or speculate). Unrealized gains or losses on such non-qualifying derivatives positions are carried on the balance sheet and reported on the income statement (under other income) of each reporting period. 7

10 risky transaction is recognized (until then, these gains or losses are only reported on the balance sheet as unrealized gains or losses). For instance, if a firm hedges the selling price of half its production a year ahead of time, then half the sales reported in the current quarter reflect last year s price and half reflect this quarter s spot price. Therefore, as a result of cash-flow hedge accounting, the sales or costs of firms that hedge are an amalgam of their past hedging decisions: Products (supplies) delivered (received) in the current quarter might have been hedged one, two, or many quarters in the past or not at all. In other words, we can decompose sales as follows: n 3 Q 3 + F 6 Q 6 + F 9 Q F n Sales0 = P0 Q0 + F... Q (1) + n n where P 0 is the spot price observed in the most recent quarter, Q 0 is the quantity of product sold at the spot price, t F + t is the t-month futures price observed t months earlier, and t Q + t quantity of product hedged t months earlier. A similar decomposition can be done for costs. is the The procedure to detect the hedging footprint from this decomposition follows directly: We can uncover a firm s hedging policy by regressing its sales (or costs) against suitably lagged futures prices and the recent spot price. The resulting regression coefficients ( Q + 3 3,, n Q + n ) represent the portion and maturity of the firm s production that is typically hedged. By confining the regression coefficients to the zero-one range ( Q 0, t Q + t + + and requiring that the coefficients sum to one ( Q 0 + percentage of a firm s risk exposure that is typically hedged: n n Q + t Q + t,, t Q + t,, n Q + [0,1]) n = 1), we can compute the Hedge Intensity HI Q t Q + t + + n n Q + [0,1] () and the percentage that is left unhedged: Q 0 = 1 HI. We can also estimate a firm s hedge maturity (in months) as the time-weighted average of the regression coefficients: Hedge Maturity HM t + n ( Q Q t t Q n n) [0, n] (3) HI 8

11 Finally, we should note that by stripping out much hopefully, most of the corporate hedging activity from firms accounting data, our approach allows us to re-estimate the value of corporate risk management presented in MacKay and Moeller (007), where only the 3-monthlagged 3-month futures prices was introduced as a remedy to partly account for corporate hedging activity reflected in the data (versus the full complement of futures prices used here). Bringing together the hedge intensities and hedge maturities for sales and costs allows us to examine both sides of the operation in ways that have not been done previously: Is the hedge intensity the same for sales and costs? If it is different, why? Is the hedge maturity the same for sales and costs? If it is different, why? What is the scope for natural hedging (the degree to which sales and costs correlate)? Etc. In other words, our proposed approach allows us to explore several related dimensions of hedging policy and determinants of corporate risk management. We should point out that because derivatives positions that do not qualify as cash-flow hedges are subject to different accounting rules, namely, normal GAAP, our proposed approach helps to sort true hedging activity from other uses of derivatives, such as speculation and market timing, also known as selective hedging (see Stulz, 1996). However, because cash-flow hedge accounting departs from GAAP, the Financial Accounting Standards Board (FASB) is conservative in determining what derivatives positions qualify as cash-flow hedges. Our approach may therefore understate actual corporate hedging activity for this reason. Another reason is that fair-value hedges do not qualify for the same special accounting treatment as cashflow hedges and therefore do not leave a similar footprint (all fair-value hedge positions are marked to market each reporting period, regardless of when they were initiated or maturity). In other words, the proposed approach can detect hedging related to the income statement but not hedging related to the balance sheet. We see in this more of a strength than a shortcoming: By focusing only on cash-flow hedges, our approach actually offers a tighter link to theory because all extant models of hedging set out to explain why firms would rationally seek to stabilize cash flow, not why they might engage in fair-value hedges or speculative hedges. 9

12 B. Empirical specification and econometrics We propose to use constrained nonlinear Generalized Method-of-Moments (GMM) methods to estimate the following simultaneous-equation specification, which contains equations for sales, costs, and the derived output-supply and input-demand equations: = a p + Y0 [ bp p0 + c p p0 ] + Y 3 [ bp p 3 + c p p 3] Y 4 [ bp p 4 + c p p 4] + controls + s (4) Sales μ ~ = aw + X 0 [ bww0 + cww0 ] + X 3 [ bww 3 + cww 3] X 4 [ bww 4 + cww 4] + controls + c (5) Costs μ ~ = Y0 [ bp + c p p0] + Y 3 [ bp + c p p 3] Y 4 [ bp + c p p 4] + y μ ~ y (6) = X 0 [ bw + cww0 ] + X 3 [ bw + cww 3] X 4 [ bw + cww 4] + x (7) x μ ~ and Where p 0 and w 0 denote current-quarter averages of daily output and input spot prices, w are lagged month-t futures prices, y is Sales/p + 3 0, and x = Costs/w 0. The hedge rates ( Y, t Y, Y, Y, Y, Y and X, X, X, X, X, X p t 3 ) are the weights the estimation assigns to the lagged month-t futures prices versus current spot prices. Consistent with the framework developed in MacKay and Moeller (007), we include squared values of the prices to capture the curvature of the sales and cost functions and the value of corporate risk management. The number of contracts (from 3 to 4 months) we propose to use reflects the availability of contracts over the empirical time period, which varies from as few as ten months early on to as many as seven years in recent years (however, liquidity is highly skewed toward the nearest maturities). 8 Control variables are as in MacKay and Moeller (007). All firm-specific variables (sales, costs, output, input, and some control variables) are normalized by dividing through by net property, plant, and equipment (because of the way accounting for derivatives works, dividing by total assets is inappropriate since unrealized gains and losses on derivatives are reflected in current asset accounts and shareholders equity). Finally, all variables are logged. 8 In order to preserve the full range of the empirical sample period (1985 to 004), we use adjacent contracts to interpolate or extrapolate prices for the 1-month, 18-month, and 4-month contracts when these are unavailable. 10

13 We estimate the model in first differences to alleviate three econometric problems: 1) the energy prices series are probably non-stationary (especially in recent years), ) even though the lagged futures prices are not contemporaneous they may still be highly correlated, which could bias the standard errors because of multicollinearity, and 3) first-differencing is equivalent to controlling for firm-fixed effects and thus addresses simultaneity bias caused by endogeneity and coefficient bias caused by omitted correlated variables. (Note: Current results are still in levels.) The measures of corporate hedging activity (intensity, maturity, and value of CRM) are estimated both for the entire sample (to identify the best specification) and at the firm level. The latter estimates can then be used to explore the determinants of corporate risk management. II. Data We implement our analysis on a sample of oil refiners. Several reasons make the oil refining industry a good candidate for study. First, energy prices swing widely, and this variation contributes to the empirical fit of the model. This is particularly important here because we use quarterly accounts rather than stock returns. Second, oil refining is a well-defined operation, with highly competitive commodity markets on both the input and output sides of the business where crude oil is the main input, and heating oil and unleaded gasoline are the main outputs. Finally, the petroleum and oil refining industries have been studied in several prior papers (e.g., Gibson and Schwartz (1990), Litzenberger and Rabinowitz (1995), Schwartz (1997), Haushalter (000), Brown and Toft (00), Borenstein and Shepard (00), Haushalter, Heron, and Lie (00)). Our firm-level data for oil refiners (SIC 911) are from the merged CRSP-COMPUSTAT quarterly data set maintained by Wharton Research Data Services (WRDS). The main variables we use are: sales (data item #), costs (cost of goods sold, item #30, minus depreciation and amortization, item #5), book value of assets (item #44), net property, plant, and equipment (item #4), and working capital (current assets, item #40, minus current liabilities, item #49). 11

14 Below we validate our approach by regressing firm value on our risk management values. Our proxy for firm value is Tobin s q, which we measure as the market-to-book value of assets. We obtain the market value of assets by replacing the book value of equity by its market value (number of common shares outstanding, item #61, times the quarter-end share price, item #14). Following Allayanis and Weston (001), these cross-sectional regressions include a number of other variables as controls, namely, total debt (short-term debt, item #45, plus long-term debt, item #51), capital expenditures (item #90), dividends (common dividends, item #0, plus preferred dividends, item #4), and research and development (item #4), all divided by the lagged book value of assets (item #44). Some of these control variables have very poor coverage. For instance, research and development is missing for over 75% of the sample. We therefore set missing control variable observations to the industry-year mean to avoid serious sample attrition. Some of the quarterly data are actually semiannual or annual (COMPUSTAT codes these as.s and.a). We identify and treat such cases as follows. For flow variables (sales, costs, etc.), we use the semiannual observation divided by two and the annual observation divided by four. For stock variables (assets, inventories, etc.), we use the most recent observations available. We also try simply deleting such observations. This causes the sample to drop from 34 to 31 firms but does not alter our conclusions. We use annual COMPUSTAT business segment data to construct two additional control variables, namely, vertical integration and diversification. Vertical integration measures a firm s involvement in so-called upstream industries (production and exploration) and downstream industries (chemicals, distribution, marketing, etc.) relative to its core business (oil refining). 9 Diversification measures a firm s involvement in industries unrelated to refining. Using segment 9 Specifically, we classify the following segments as upstream industries: two-digit SIC 13 (exploration and production of crude and natural gas) and four-digit SIC 461 (crude oil pipelines) and 679 (oil and gas royalties and leases). We classify the following segments as downstream industries: two-digit SIC 8 (chemicals), 30 (plastic products), 46 (pipelines), 49 (natural gas transmission and distribution), 51 (wholesale petroleum-based products distribution), 87 (engineering, management, and consulting services), and four-digit SIC 3533 (oil and gas field machinery), 5541 (gasoline stations), 5984 (propane marketing), and 7549 (fast lube operations). 1

15 data, we measure vertical integration as one minus the Herfindahl of a firm s refining-related segments and diversification as one minus the Herfindahl of its nonrefining-related segments. Input and output prices are constructed as follows. We obtain daily settlement prices, volume, and open interest for all NYMEX-traded futures contracts on light crude oil, heating oil, and unleaded gasoline from Thompson Financial s Datastream International. Beginning in March 1985, delivery months for all three commodities have been available for every month of the year going out several months. These commodities represent the main outputs (heating oil and unleaded gasoline) and input (crude oil) for the oil refining industry (SIC 911). 10 To simplify our analysis, we exploit a useful feature of the oil refining process, namely, that these inputs and outputs are roughly consumed and produced in the following proportions: three barrels of crude oil yield approximately two barrels of unleaded gasoline plus one barrel of heating oil. The price difference between contracts held in these proportions (3::1) is known as the crack spread, and the contracts traded on NYMEX reflect this ratio (NYMEX (000)). We combine the prices of heating oil and unleaded gasoline into a single output price, weighting each price according to the crack spread ratio. The resulting price therefore represents two-thirds of the gasoline price plus one-third of the heating oil price. Tracking one output price instead of two makes the analysis more tractable. Figure 1 shows input and output prices and the crack spread from March 1985 to June 004. Because our panel runs from March 1985 through June 004 we need a deflator to make firm variables and prices comparable across time. We use the monthly consumer price index #SA0L1E (All items less food and energy) produced by the U.S. Bureau of Labor Statistics (BLS). We use a deflator that excludes energy prices because we want to remove the effect of general inflation without removing the effect of energy price changes. We scale the deflator and the input and output prices relative to their March 1985 levels, the first month of our panel. 10 Following Litzenberger and Rabinowitz (1995), we use the nearest-month futures contract to construct our time series of spot prices. Datastream uses the previous business day s settlement price for holidays (when reported volume is zero). We therefore exclude these and any other zero-volume daily observations. 13

16 Because our firm-level data are quarterly, the next step is to convert our input and output price series from daily to quarterly series. We consider three weighting schemes to aggregate the daily data into quarterly observations. First, we use a volume-weighted average to guard against stale data and to avoid giving equal importance to prices associated with unusually low or high trade volume. Second, as a variant on this scheme, we also try weighting prices by the daily level of open interest. Third, and simplest, we equally weight the daily observations. Although the three weighting schemes produce similar results, we retain the volume-weighted scheme because it seems most suitable. Weighting by volume also accounts for times when trade volume in the futures contracts differs substantially from the level of trade in the nearest-month (spot) contract. We match the firm-level quarterly data to the price data by mapping fiscal year-quarters to the appropriate calendar year-quarters. Because fiscal year-ends can occur in any month of the year, we match firm data to quarterly price averages constructed for each month of the year. III. Summary Statistics Figure 1 shows nominal quarterly spot and three-month futures prices from March 1985 to June 004. The graph shows that input and output prices vary widely, fluctuating between roughly 13 and 48 dollars per barrel and that the difference between the output and input price the crack spread understandably trades in a much smaller range of to 10 dollars. Although the magnitude of the crack spread is much smaller than the output and input price, each penny change in the spread translates into millions of dollars for the average oil refiner. A back-of-theenvelope calculation shows that for the mean firm in our sample, a one-cent change in the crack spread causes a $.5 million change in quarterly operating cash flow in 1985 dollars. Insert Figure 1 around here. Table I shows summary statistics for the 78 quarterly spot and futures energy prices in our sample period. The mean nominal output and input spot prices are 6.43 and 1.87 dollars per barrel while the mean crack spread is 4.56 dollars. Figure 1 and Table I show that the futures prices are generally below the spot prices, indicating backwardation both in the price of crude oil 14

17 (as in Litzenberger and Robonowitz (1995)) and in the output prices (gasoline and heating oil). Input and output prices are highly correlated (0.99 for both spot and futures), as are spot and futures prices (0.99 for both input prices and output prices but only 0.71 for the crack spread). Insert Table I around here. Figure shows aggregate statistics for the U.S. refining industry. These include annual production and consumption of refined petroleum products and refinery capacity utilization rates. Using a price index of refined petroleum products from the Bureau of Labor Statistics from 1977 to 003, we estimate that the price elasticity of demand (consumption) is -10%. Insert Figure around here. Table II, Panel A reports summary statistics on the operating characteristics of our sample of 34 oil refiners obtained from quarterly COMPUSTAT data. The data show that oil refining is a large-scale, capital-intensive activity (mean assets near $1 billion, net plant, property, and equipment nearly 50% of assets, capital expenditures nearly 5% of assets per year), that operates on thin margins (mean operating cash flow is 5.4%), and is characterized by relatively low market-to-book value of assets (mean Tobin s q is 1.6). Insert Table II around here. We later use the hedging activity reflected in the data to estimate hedge rates. To put those estimates into perspective, Table II, Panel B reports summary statistics on derivatives usage by our 34 sample firms. Following previous studies (e.g., Geczy, Minton, and Schrand (1997), Allayanis and Weston (001)), we conduct a systematic search of our sample firms annual reports for all discussions of risk management policy and practice. Specifically, we search for the words risk, hedge, forward, futures, derivative, swap, option, and index. We find that hedging policies and derivatives usage are barely mentioned prior to Discussion of these subjects 15

18 has become more detailed as disclosure requirements have increased. 11 From this search we construct two measures of derivatives usage, one measuring the level of derivatives usage, the other measuring the type of derivatives used. 1 The first measure classifies firms according to whether they rarely hedge, sometimes hedge, or usually hedge. We find that twenty oil refiners usually hedge while seven sometimes hedge and seven others rarely hedge. All sample firms report using derivatives except for one, Imperial Oil, which explicitly states that its policy is not to use derivatives. Two firms, ConocoPhillips and ExxonMobil, report some derivatives usage even though they have a stated policy of remaining exposed to price fluctuations or relying on diversification to manage risk. 13 The second measure sorts firms based on the type of derivatives they use. Firms that use energy-related derivatives we classify as operating. Firms that use financial derivatives (e.g. interest rates, foreign exchange rates) we classify as financial. We find that eight firms use operating derivatives only, four firms use financial derivatives only, and firms use both. Intersecting these two measures of derivatives usage, we find that the largest subgroup in our sample consists of the 15 firms that usually hedge and that use both operating and financial derivatives. It is important to recognize that the FASB rules regarding the treatment of derivatives apply to conventional definitions of derivatives (futures, options, swaps) and do not necessarily include nonderivatives-based hedges such as long-term arrangements refiners make with clients. For instance, in its 00 annual report Amoco explains that it enters into fixed-price agreements for marketing purposes with its clients and may use derivatives to offset these contracts if the 11 The Financial Accounting Standards Board (FASB) issued a series of statements intended to improve the transparency of derivatives usage. A review of these statements is available from the authors. 1 Although we would prefer to use a continuous measure of derivatives usage, the data disclosed by our sample firms on their derivatives positions are too sporadic and inconsistent to construct a meaningful continuous measure. We therefore fall back on the categorical variables presented here and commonly used in prior studies. 13 The results of our annual report search, including the lack of a meaningful continuous measure, that only one firm does not use derivatives, and that a few firms indicate they rely on integration to manage risk, are independently confirmed by the U.S. Energy Information Administration in Derivatives and Risk Management in the Petroleum, Natural Gas and Electricity Industries. 16

19 associated cost basis has not been hedged or otherwise fixed. This example points to a limitation of derivatives usage as a proxy for risk management. The hedge rates we present later help to overcome this limitation of the derivatives usage measures. Recent work on selective hedging (e.g., Brown, Crabb, and Haushalter (003), Adam and Fernando (005)) illustrates another way in which observed (or stated) use of derivatives does not tell the whole risk management story. IV. Regression Model Estimation A. Econometric Approach Table III presents General Method of Moments (GMM) coefficient estimates for the set of simultaneous equations represented by expressions (4) to (7) for a pooled sample of oil refiners. These equation systems represent the revenue and cost functions and their associated derived output-supply and input-demand equations. The dependent variables for these equations are, respectively sales, costs, output quantity (sales divided by output price), and input quantity (costs divided by input price). The table presents layered versions of the model to examine the effect of including progressively longer-dated lagged futures prices (Models 1 to 8). No separate column appears for the input and output equations because the sales and costs equations already reflect all the model coefficients. We include the input and output equations in the estimation because the added structure reflects the firm s first-order conditions and the state of its product and factor markets. Including these equations also improve the efficiency of the coefficient estimates. In contrast to Ordinary Least Squares (OLS), GMM allows for simultaneity among the dependent variables by incorporating the correlation of residuals across the four equations. This improves the efficiency and consistency of the estimates. As an instrumental variable estimation method, GMM mitigates simultaneity bias caused by endogenous explanatory variables by using predicted (instrumented) values rather than realized values of the endogenous variables. We instrument the endogenous variables (all variables except prices) by the first to fourth powers of the spot, futures, and lagged futures prices for inputs and outputs (40 instruments). 17

20 We use Hansen s (198) J-statistic to jointly test whether the model is well specified and the instruments are valid. We also use the J-statistic to assess the gain or loss in overall fit across model specifications. For every model in Table III we find J-statistics significantly different than zero, which represents a rejection of the overidentifying restrictions and implies that the model is not fully specified, the instruments are correlated with the residuals, or both. Comparing models one and two, we find that adding even a single lagged futures price (e.g., the 3-month contract) substantially lowers the J-statistic, suggesting that, as Leamer (1983) shows, large-sample specification tests are sensitive to even small departures from the true model. However, even in our preferred specification (Model 8), where the J-statistics are lowest, the over-identifying restrictions are still rejected, suggesting that some simultaneity bias remains. The chosen instrument set reflects a best-efforts balance between validity (instruments uncorrelated with residuals) and relevance (instruments correlated with the endogenous variables). Although we address heteroskedasticity by normalizing the firm-level variables by the lagged book value of assets, this might still pose a problem. Additionally, sales, costs, and prices all exhibit autocorrelation (see Tables I and II). Heteroskedasticity and autocorrelation can bias the standard errors and over- or understate both the statistical significance of the variables and the precision of our estimates of the value of risk management. We therefore use a first-order autocorrelated Newey-West (1987) procedure to correct for these econometric problems. Our regressions also include unreported fiscal-quarter dummy variables to adjust for seasonality. Insert Table III around here. Note to the reviewer or reader: The following sections are being rewritten. For now, please jump to the tables. 18

21 50 40 $/barrel Output Spot (Heating Oil & Gasoline) Input Spot (Crude Oil) Output 3M Futures Input 3M Futures Spot Crack Spread 3M Crack Spread Figure 1. Quarterly energy prices from March 1985 to June 004. Quarterly energy spot (nearest-month) and three-month futures prices constructed from daily NYMEX-traded futures contracts on light-crude oil, heating oil, and unleaded gasoline from Datastream. We construct quarterly price series from trade-volume weighted averages of daily closing prices. The output price, p, is one-third of the price of heating oil plus two-thirds of the price of unleaded gasoline. The input price, w, is the price of light crude oil. The crack spread, s, is the difference between the output price and the input price.

22 thousands of barrels (in logs) 100 % Consumption Production Capacity Utilization Figure. Annual oil refining statistics 1977 to 003. Annual data on U.S. production and consumption of refined petroleum products and refinery capacity utilization. Based on a refined petroleum product price index from the Bureau of Labor Statistics, the estimated price elasticity of demand (consumption) is -10%. Source: U. S. Department of Energy (Energy Information Administration). 60

23 Output Futures Prices (Gasoline & Heating Oil) Input Futures Prices (Light Crude Oil) Figure 3a. Level of output and input futures by maturity Q to Q 004.

24 30,000 Trade Volume of Output Futures (Gasoline & Heating Oil) 5,000 0,000 15,000 10,000 5,000 0 Trade Volume of Input Futures (Light Crude Oil) 100, , , ,0000 0, Figure 3b. Trade Volume for output and input futures by maturity Q to Q 004. Open Interest in Output Futures (Gasoline & Heating Oil)

25 50,000 40,000 30,000 0,000 10,000 0 Open Interest in Input Futures (Light Crude Oil) 180, , , , , , Figure 3c. Open interest for output and input futures by maturity Q to Q 004.

26 Table I Summary Statistics: Quarterly Energy Prices Quarterly energy spot (nearest-month) and three-month futures prices constructed from daily NYMEX-traded futures contracts on light crude oil, heating oil, and unleaded gasoline from Datastream for March 1985 through June 004. We construct quarterly price series from trade-volume weighted averages of daily closing prices. The output price, p, is one-third of the price of heating oil plus two-thirds of the price of unleaded gasoline. The input price, w, is the price of light crude oil. The crack spread, s, is the difference between the output price and the input price. Spot (Nearest-Month) Prices 3-Month Futures Prices Output Price, p Input Price, w Crack Spread, s Output Price, p Input Price, w Crack Spread, s Observations (quarters) Mean Median Standard Deviation Skewness Kurtosis Minimum Maximum Correlations (p & w) Correlations (Spot & 3M) ARMA (p,q) (1,1) (1,1) (1,1) (1,1) (1,1) (1,1) 1 st -order autocorrelation

27 Table II Summary Statistics: Quarterly Firm Operating Data and Derivatives Usage Panel A shows summary statistics for a sample of 34 oil refining firms (SIC 911) from 1985 to 004. Quarterly COMPUSTAT data definitions: sales (item #), costs (cost of goods sold, item #30, minus depreciation and amortization, item #5), book value of assets (item #44), fixed-capital (net property, plant, and equipment, item #4), working capital (current assets, item #40, minus current liabilities, item #49), Tobin s q (market-to-book value of assets, where the market value of assets is obtained by replacing the book value of equity by its market value (common shares outstanding, item #61, times the quarter-end share price, item #14)), total debt (short-term debt, item #45, plus long-term debt, item #51), capital expenditures (item #90), dividends (common dividends, item #0, plus preferred dividends, item #4), and research and development (item #4). All normalized variables are divided by the lagged book value of assets. Some of these variables have very poor coverage so we set missing values of control variables (namely, total debt, capital expenditures, dividends, and research and development) to the industryyear mean to mitigate sample attrition. Vertical integration measures a firm s involvement in upstream industries (production and exploration) and downstream industries (chemicals, distribution, and marketing) relative to oil refining. Diversification measures its involvement in industries unrelated to oil refining. Using COMPUSTAT business-segment data, we measure vertical integration (diversification) as one minus the Herfindahl of a firm s oilrelated (unrelated) business segments. Panel B shows derivatives usage derived from annual reports, classified by hedging level (rarely, sometimes, and usually) and type of risks hedged (operational, financial, and both). Panel A. Quarterly Firm Operating Data Mean Median St. Dev. Within Firm Variation Min Max 1 st Order Autocorrelation Sales (in million $) 3,49 1,081 5,330 13%.16 35,769 83% Costs (in million $), ,955 16% -.4 9,857 81% Size (in million $) 1,088 3,539 17,033 3% ,916 95% Operating Cash Flow / Assets 5.40% 5.47%.34% 56% -9.70% 4.06% 40% Fixed Capital / Assets 48.55% 49.4% 7.66% 40% 17.6% 64.6% 83% Working Capital / Assets 3.84% 4.7% 11.69% 45% -167% 5.6% 77% Tobin s q % % Total Debt / Assets 6.48% 4.37% 13.40% 31% 0.00% 9.46% 85% Capital Expenditures / Assets 5.13% 4.1% 3.89% 83% 0.00% 7.49% 13% Dividends / Assets 0.06% 0.01% 0.19% 61% 0.00%.01% 86% R&D / Assets 0.1% 0.1% 0.04% 67% 0.00% 0.8% 91% Vertical Integration 6.55% 7.95% 0.39% 36% 0.00% 74.58% 89% Diversification 5.45% 0.00% 11.65% 5% 0.00% 54.89% 87% Observations (firm-quarters),145 Panel B. Derivatives Usage Hedging Level Risks Hedged Rarely Sometimes Usually Total Operational Financial Both Total