IS TEMPERATURE THE SINGLE MOST IMPORTANT FACTOR IN how gas storage is used? Or are other variables involved? Can we answer these questions - and verify the results?
Since FERC Order 636, natural gas storage has grown into a high-profile asset in the industry. As a result, the industry has responded by changing the way it uses this storage. The exact nature of this adjustment is not apparent at a glance; one must first analyze industry data.
For example, in its 1995 study of natural gas storage, The Value of Underground Storage in Today's Natural Gas Industry, the Energy Information Administration showed storage operators were withdrawing more gas from storage in the summer and injecting more in the winter. Many in the industry were surprised by the degree that this was occurring.
This change in the role and use of storage relates in part to its use as a system-balancing tool. Before Order 636, when pipelines owned the gas flowing in their systems, they could use that gas to achieve system balancing. Now, it is up to individual shippers to acquire the right amounts of gas and service to sustain system reliability. Under competition, these shippers also can take advantage of arbitrage opportunities that surface in a price-volatile commodity market.
Our analysis of storage suggests certain characteristics about its use and provides us with some interesting and useful results. First, it appears that the industry's operating behavior follows almost the same "script" from year to year. Second, while storage use has grown more flexible since Order 636, it still reflects risk-averse behavior or an inability to respond to certain market signals.
Yes, storage withdrawals are strongly linked to temperature changes - in the way we expected. And yes, withdrawals are related to price changes - but opposite from the way we expected. The information presented here summarizes how the industry is responding to restructuring, including reactions to storage levels and withdrawals, temperature and price during the last several years.
Identifying Three Variables
Today, information on storage, including holding and delivery capacities and construction of new facilities, is widely sought and reported. Data on working gas inventories rank alongside spot and futures prices throughout the year, but particularly during heating seasons. In fact, storage data, along with short-run weather data, are probably the most watched numbers in the industry.
Leading into the heating season, interest in weekly inventory and withdrawal numbers intensifies, because storage provides the backup to pipeline supply and is the key source to satisfy excess demand.
Another variable that might affect withdrawals is the price of gas. For instance, if prices increase, then one might expect larger withdrawals, as holders of storage inventory take advantage of market opportunities. And if prices are falling, then one might expect smaller withdrawals and companies might even add to their inventories by taking advantage of low-priced gas.
Finally, inventory levels with respect to anticipated or expected demand also might affect withdrawals. That is, how much gas is on hand vs. how much is expected to be needed. What one might expect is that, as inventory relative to expected demand decreases, withdrawals would decrease. If inventories are high relative to expected demand, then withdrawals would tend to be larger.
In 1993, the American Gas Association began publishing estimates of weekly working gas inventories held in storage facilities in three regions: Producing, Consuming East, and Consuming West (see Figure 1). Since A.G.A. storage data are the only weekly series widely available, they are relied upon heavily for short-term business decisions and greatly influence market behavior.
Data from the National Weather Service coupled with the withdrawal and inventory data from the A.G.A., covering three full heating seasons and part of a fourth, makes it feasible for us to conduct some basic statistical analysis of the relationship between temperature and withdrawals of gas from storage.
We analyzed this data to understand the relationship between storage, withdrawals and price, temperature, and inventory levels relative to expected demand and to develop a way to estimate future storage withdrawals and real-time inventory levels. Our goal was to define a relationship that can be interpreted meaningfully using linear regression techniques.
Our next task, therefore, was to refine further and to quantify the variables that we used to represent variations in temperature, prices, and inventory levels relative to expected demand.
Isolating Temperature
The seasonality of withdrawals should not surprise anyone: More gas is used in the winter, including more from storage, to satisfy seasonal demand (see Figure 2). Weekly average temperatures from four major gas-consuming cities within the Consuming East region illustrated this seasonality (see Figure 3).fn1 By comparing Figures 2 and 3, we saw that they were almost mirror images of one another; withdrawals increased as temperatures decreased.
However, when we looked closely at these two figures, there were some interesting and surprising elements. For example, for many heating season weeks, storage withdrawals remained close to the same value from year to year. On the other hand, toward the middle of the heating season, around weeks 9 through 11, we found great variability in both temperatures and withdrawals. Withdrawals values toward the end of the heating season tended to fall somewhat lower than at the beginning of the heating season.
The "shapes" of the distributions of both withdrawals and temperatures appeared fairly symmetrical. However, upon closer examination, the distribution of withdrawals seemed somewhat flat starting out the heating season. Then, as the heating season progressed to its midpoint, withdrawals appeared to spike; but from mid-season to its end, withdrawals showed a more shallow downslope. Yet the upward and downward slopes of the temperature distribution appeared to have about the same degree of "steepness."
These observations suggested that storage was used in a similar manner year-to-year at the start of the heating season in a way that is somewhat independent of temperature. The steep ramp-up of storage withdrawals in those months reflected the fact that, early in the heating season storage facilities are relatively full and carry higher operating pressures, and thus greater deliverability. Perhaps the flat shape early in the season reflected a tendency to hold on to stored gas early in the season, to have plenty for that "coldest day."
At the end of the heating season, withdrawals tended to fall to lower levels than at the beginning of the heating season, reflecting the fact that inventories have been drawn down, reservoir pressures are low and resulting deliverability is reduced.
This year-to-year seasonal pattern of storage use suggested a link to standard industry practice. For our analysis, we needed to remove this seasonality from the data. By "deseasonalizing" the data, we removed the variation in the data that could be attributed to standard industry practice. We did this to isolate the effect of temperature on changes in withdrawals, independent of the seasonal operating practices of the industry.
There are several ways to "deseasonalize" the data. We chose to take the difference in both withdrawals and temperatures between current and past seasons. To derive the withdrawal differences, we took the withdrawals of any given week in the current heating season and subtracted the withdrawals in the same week of the previous season. To derive the temperature differences, we took the average temperature in any given week of the previous heating season and subtracted the average temperature in the same week of the current heating season. Positive differences meant greater withdrawals and colder temperatures, respectively, in current years over previous years. Figure 4 reveals how strong the linear relationship is between these two series.
When we regressed our deseasonalized withdrawal information on our deseasonalized temperature information, we found that temperature explained more than 75 percent of the variability in our withdrawal series.
Constructing the Price Variable
To construct our price variable, we repeated the technique of taking differences. But this time we compared the price of gas during the current heating season to what it might have been when purchased for storage. A large part of working gas inventories withdrawn during a given heating season are injected into storage during the immediately preceding non-heating season, i.e., the months of May through October. In fact, over time we expect an increasing proportion of the gas injected during a non-heating season to be used during the subsequent heating season, as a mark of the industry's increasingly efficient use of inventories in a competitive environment.
We computed a weighted-average price for gas during the non-heating season. For this computation, we used the Transco Zone 65 price series,[fn2] weighted by the ratio of weekly net injections reported by A.G.A. for the 29 weeks not included in our definition of the heating season, to the total of net injections during those 29 weeks. This gave us an estimate of the average cost of gas during the injection season. Next, for each week in a particular heating season, we subtracted this weighted average injection season price from each of the Transco Zone 6/CNG weekly prices.
Positive price differences indicated heating season spot gas prices are higher than what might have been paid for gas injected into storage. Negative differences indicated heating season spot prices are less than an average of non-heating season (or injection season) prices. Negative values also implied that costs of transporting and storing the gas were not recovered (see Figure 5).
This price series illustrated well the price shocks the industry has undergone. The large 1995-96 heating season figures, for instance, show the value of stored gas increased significantly during the heating season. The negative 1994-95 figures indicated the value of stored gas had declined significantly during the heating season. The results here would prove even more striking if we had subtracted the cost of transporting the gas to the Transco Zone 6 market area and then storing the gas until the heating season. This cost could easily exceed $1.25 per Mcf.
Inventory Levels vs. Expected Demand
Since weekly demand data was not available, we decided to approximate the figures based on the assumption that gas demand and temperature are inversely related.
A measure of expected demand at one temperature relative to another would simply be the ratio of two temperatures. For example, the expected demand at 30 degrees Fahrenheit relative to 60 degrees would be the ratio of 60 to 30, or 2. We constructed a time series of relative expected demand by computing average temperatures for the heating season weeks based on the last 13 years. We then selected the highest average temperature and divided it by each of the 13-year weekly average temperatures. This step yielded a series of weekly relative demand indices, which were assigned to 1994.
Next, this series of weekly indices was inflated at a compound rate of 2.5 percent per year for each of the succeeding three years of 1995, 1996 and 1997. This was done to account for the average annual growth in demand of natural gas during that period. Finally, for each week, we computed the ratio of the level of gas in storage divided by the expected demand index for the following week. The resulting variable is a proxy for the ratio of inventory levels to expected demand.
Estimating a Regression Equation
We used standard regression to estimate the linear relationships between withdrawals and three variables: temperature, price and demand. Our model includes four numerical constants - a, b, c and d - and took this form:
Year-to-year differences in withdrawals for each week = a + b 3 (year-to-year differences in temperature for each week) + c 3 (price difference for each week, heating-season minus non-heating season) + d 3 (inventory to expected demand index for each week) + a random error term for each week.
We estimated values for the equation's constants, using ordinary least squares (OLS) and assumed that the OLS assumptions would be satisfied.
When we regressed withdrawals on the variables described above, we obtained an R-squared value, which shows how well the model fits to the data points, of 0.78. The regression coefficient for the variable for differences in weekly average temperatures between consecutive heating seasons was a positive 4.1. For each change of 1 degree Fahrenheit in the difference in average temperatures in a given heating season week between two consecutive years, the difference in withdrawals between years is 4.1 Bcf.
The coefficients for our two other variables were both negative. The two variables were the weekly spot prices during the heating season minus the weighted-average price of gas during the preceding non-heating season, and the proxy variable for the ratio of inventory levels to expected demand. In addition, the T-value for the proxy variable for the inventory to expected demand ratio had an absolute value less than 1.6, which indicated that the estimated regression coefficient for this variable is not statistically different from zero. Therefore, we dropped this variable from consideration and focused on the modified model specified below.
Our modified regression equation:
Year-to-year differences in withdrawals for each week = a + b 3 (year-to-year differences in temperature for each week) + c 3 (price differences for each week) + a random error term for each week.
When we ran this regression, we obtained an R-squared value of 0.78. Thus, the two proposed independent variables "captured," or explained, about 78 percent of the variability in year-to-year differences in withdrawals. In other words, changes in amounts of gas withdrawn from storage during the same week of consecutive heating seasons, at least in A.G.A.'s Consuming East region, are strongly dependent on differences in weekly average temperatures (associated with our four "observation" cities of New York, Pittsburgh, Chicago and Kansas City) for the same week of consecutive heating seasons, and on the difference in price between weekly heating season spot prices and an average non-heating season spot price. Our regression results yielded the following estimation equation:
Withdrawal differences = 6.3 + 4.1 3 (Temperature differences) - 4.0 3 (Price differences). The T-values for the three estimated quantities are as follows:
Constant or Intercept: 1.97, coefficient of temperature differences: 14.41, coefficient of price differences: -2.86.
Interpreting the Results
The negative sign of the price difference coefficient proved to be opposite to what we had expected. We had reasoned that if heating season spot prices exceeded some estimate of the cost of stored gas, then the owners of that gas would tend to withdraw more from storage. But just the opposite was true: As the difference between heating season spot prices minus injection season average prices increased, withdrawals tended to fall.
This result may reflect our having estimated one equation when at least two equations estimated jointly would have been better. One equation would represent the relationship between price and supplies of gas to market, which includes not only withdrawals but also other available supplies of gas, such as imports and domestic production. These variables would represent gas that was previously contracted for but which the holder would be willing to release. It also would represent additional gas coming from Canada and domestic production that had not been contracted out. But actual measurements or good proxies are not available for these variables. The second equation would be similar to the equation estimated here.
Another interpretation of the negative sign on price is that owners of stored gas in this industry have risk-averse tendencies. As the heating season progressed, if and when spot prices rose, owners of stored gas expected prices either to remain at the higher levels or continue to rise. Because they were risk-averse, they tended to hold onto more of their stored gas. They feared that using their stored gas immediately could expose them to even higher prices later. This apparent conclusion in one sense runs counter to the popular notion expressed lately that storage use has become a lot more flexible since Order 636. Yet while operational flexibility has increased, it may not be motivated by price, or at least not by price as we have represented it.
This theory led to another possible interpretation. It is possible that much of the stored gas in the Consuming East region was owned by local distribution companies, who, because of current regulatory structure, had little incentive to, or could not, predictably respond to price signals.
Figure 6 is a plot of the residuals from our regression analysis, that simply represents the variation in withdrawal differences not explained by our regression equation. We hoped that they would scatter around the zero line like a random error term for each week. But the figure showed that the residuals tended to increase as we progressed through the heating season, indicating the increased uncertainty associated storage withdrawals as the end of the heating season approaches.fn3 Also, the same week in different years gave us the highest and the lowest residuals.
The largest residual value (the observed value was greater than the expected value) occurred in the week ending Feb. 9, 1996, when prices in Transco Zone 6 were more than $10 per million Btu. Thus, although high prices on average lead to conservative behavior, exceptionally high prices may encourage the release of gas to the market. The low withdrawals for the week ending Feb. 7, 1997 were difficult to understand since, although prices were not particularly high or low, the storage level relative to expected demand statistic was 46 percent higher on Feb. 7, 1997 than on Feb. 9, 1996. This observation points again to the uncertainty surrounding individual withdrawal estimates.
On average, there is a strong relationship between withdrawals and temperature, but the relationship breaks down at times. It could be that storage operators make major adjustment decisions based on a variety of factors at the beginning of February, given that by then the coldest times of the year, on average, have passed. F
John H. Herbert is a senior economist at the Energy Information Administration and an adjunct professor of statistics at the Virginia Tech Telestar Graduate Center. James M. Thompson is an industry analyst and economist at EIA. Chris Ellsworth works for ESAI Inc., a consultancy. The views expressed are those of the authors and do not necessarily represent those of the EIA, U.S. DOE or ESAI Inc. The authors thank Mike Tita for his comments and discussions.
1Weekly average temperatures are computed by adding seven daily average temperatures and dividing by seven. Daily average temperature is derived by adding daily highs and lows for all four cities and dividing by eight.
2Transco Zone 65 runs along the Transco pipeline in northeastern Louisiana, east of Henry Hub, in the supply area, where most gas is purchased in the non-heating season for storage. For the heating season, we used Transco Zone 6 price, which runs from northern Virginia to New York, in the market area. This price has been quoted only since November 1995. Prior to this, we used prices for CNG. This area represents an active, liquid cash market in the Consuming East .
3This also indicates that the constancy of the error term variance assumption for OLS was not satisfied. However, when we used a more general estimator the numerical results did not change much.
Articles found on this page are available to Internet subscribers only. For more information about obtaining a username and password, please call our Customer Service Department at 1-800-368-5001.