A Model for Predicting Natural
Gas Inventory Changes

by Matthew Millar | April 8, 2010

Print

This week I plan to flesh out the natural gas inventory model that I wrote about in my last article. Before that is introduced however, I wanted to start with a quick clarification on some comments at the end of my previous article. I had several readers ask for details on my reference to price support at the end of the last article. First of all I should note that none of what I write here ought to be taken as a recommendation to buy or sell natural gas or any other commodity/security. Aside from this, I would also say I am fairly agnostic about medium-term natural gas price, so my comment on expecting support was more "not bearish" and less "bullish." I've spent a good deal of the past 6 months being quite bearish on natural gas prices, but I recognize that:

1. Price change is seldom monotone for an extended period of time. During the period mid-February to end March, prices made new lows almost every day. The longest a low held during this period of time was ~70 hours!
2.  Average price for the forward strip last week was actually lower than it was back in September '09 when the front month was ~$2.50. Production companies certainly do Net Present Value calculations using the forward strip as their benchmark for price expectations (or for locking in prices themselves) so the fact that the forward strip was cheaper last week than it was back in September leads me to think we aren't far from where producers would change their tactics in terms of drilling and putting wells into production.

Speaking to the medium term, I think there will come a time when we have a better handle on the shale resource and then it will be more possible to predict where there might be a price inflection higher. The intricacies of land leases and M&A activity will also play a large role in timing. That being said, if shale runs into any legal issues (which I doubt), or if there is a significant overestimate of the resource (which is more likely), then the time could be approaching when we see a disruptive price spike up.

Ordinary Least Squares Natural Gas Inventory Model

In advance, a warning should be given that the following material might be a little technical for some readers' tastes. I begin to discuss applications for using the model as well as other price determinants at the end of this article. Without further ado, the model I developed is based on an OLS regression using inventory numbers as the endogenous variable and the explanatory variables given by degree days and a trend variable. The econometric results and the values for the coefficients are given in the Appendix. For a brief explanation of Ordinary Least Squares Regressions, I recommend the Wikipedia article:  http://en.wikipedia.org/wiki/Least_squares¹ .

Compare Apples to Apples

All the variables in the model are in terms of year-over-year comparisons. There are patterns to natural gas use - particularly in the industrial sector - that are fairly consistent each year. By comparing a week with the same calendar week from previous years, we "back out" the effects of these cyclical patterns (auto industry furloughs, Christmas, other holidays, refinery maintenance, etc.)

The Dependent Variable

Inventories: The variable we are trying to get to is the level of inventories. If we get a prediction for what the level of inventories ought to be, than we have some basis for judging whether the reports are bullish or bearish. More precisely, since we are looking at year-over-year comparisons, the dependent variable is d(inv) which stands for the difference between this year's inventory build/draw and the inventory build/draw from the previous year(s).

Explanatory Variables

1. Degree Days: Heating degree days and cooling degree days are the first explanatory variables. Again, since everything is based on year over year comparisons, the variables in the regression are based on differences in the hotness and coldness between the current year and previous years. The Climate Prediction Center kindly breaks out HDDs and CDDs numbers in gas weighted, population weighted terms. Thanks CPC! ftp://ftp.cpc.ncep.noaa.gov/htdocs/archives.wkthurs/
2. (Degree Days)^squared: When it is super cold or super hot, there seems to be an outsized effect on natural gas use. This is for a number of reasons. First of all, stoics may keep the heat off if it is 40 degrees outside, but most will be compelled to turn it on when temperatures drop below freezing. Second of all, severe cold usually means southern regions will have to use heat as well. These dwellings (and dwellers) are typically not as insulated and resistant to cold and will use more energy per capita than their northern brethren. Finally, "peaker" plants are almost always natural gas. The marginal demand from peak electricity loads during hot summer months fall almost exclusively to the domain of natural gas plants. All of these reasons indicate that the relationship between degree days and natural gas use is not purely linear. The regression uses the squares of the heating and cooling degree days as explanatory variables to allow for this.
3. Trend: Let's assume for a moment that the past 10 weeks of inventories have all been bearish, indicating that there is far more supply or less demand than the markets had expected. If this were the case, then the markets will have already had the chance for many weeks to "price in" the bearish fundamentals. In this circumstance, would it be reasonable to expect the market to be surprised by yet another bearish report? Obviously not; a bearish report in this context could therefore be expected to provide little new bearish ammunition. The model accounts for this by assuming there is some persistence in supply and demand imbalances. The trend variable is a proxy for this persistence, and is the weighted average of the past 5 weeks' readings, with the heaviest weighting given to the most recent week. If the previous weeks have had very bearish readings, the model therefore incorporates the likelihood of another bearish reading in the current week.

Accuracy of the Model

Over the past 3 years, this model predicts roughly 40% of the reports to within 5 BCF and 67% of reports to within 9 BCF. The R2 is 0.9 which indicates that the explanatory variables demonstrate a very high level of predictive power for the dependant variable.

Using the Model as a Tool for Price Prediction

With such a high level of accuracy, I was excited to think that the regression would be a useful tool in predicting future price action. If the market deviated significantly from what the model predicted, I expected a consistent and trade-able price reaction. I was somewhat disappointed to find that such a simple relationship does not exist. Alas, it seems that the model provides but a small piece in the puzzle of future prices. My next article will be fully dedicated to this puzzle, discussing both the use of the supply and demand model, and the more salient short-term factors such as executive expectations, applications of game theory, contractual obligations, herd behavior and other human psychological elements.

Clearly there isn't room to cover all of these topics today, so I will end this article with a teaser that challenges one to think of through these ideas, particularly the role of executive and market expectations. Go back in time exactly 4 years, 7 months, and 2 weeks. It is late summer, and crude oil has been rallying to new all-time highs throughout the summer (Figure 1.) Commercials have been net short through most of the rally and now have a large net short position. It has already been an active tropical season when the NHC names a new storm TD 12 on Tuesday, August 23, and issues alerts for Florida east coast. If you will, imagine you are a trader or an oil executive from this time, and (without looking back at a chart) think through what might happen to oil prices over the next week or two, and then over the next 4 months. Imagine how expectations, game theory, etc. might play a role in determining the future prices for oil at that time.  I love using Katrina as an example both because it defies what simple logic might predict and stands out as an extraordinary event in so many ways! The next article will pick up on this thread, and if you have any ideas you would like to share in the meantime, feel free to email me.

For the week ending April 1, there were 109 heating degree days (per the CPC gas and population weighted statistics) and my regression model predicts an inventory fill of 26 BCF.

Appendix

Dependent Variable: D1INV
Method: Least Squares
Date: 01/25/10   Time: 16:48
Sample (adjusted): 10/23/2003 1/14/2010
Included observations: 326 after adjustments
D1INV=C(1)*D1HDDS+C(2)*D1HDDS2+C(3)*D1CDDS+C(4)*D1CDDS2+C(5)*TREND(-1)+C(6)*TREND5
	Coefficient	Std. Error	t-Statistic	Prob.
C(1)	-0.580088	0.095170	-6.095279	0.0000
C(2)	-0.002792	0.000282	-9.916612	0.0000
C(3)	-0.850371	0.309677	-2.745992	0.0064
C(4)	-0.004713	0.002471	-1.907536	0.0573
C(5)	0.420884	0.053971	7.798329	0.0000
C(6)	0.362238	0.068463	5.290971	0.0000
R-squared	0.905011	Mean dependent var		0.435583
Adjusted R-squared	0.903526	S.D. dependent var		49.12565
S.E. of regression	15.25853	Akaike info criterion		8.306388
Sum squared resid	74503.27	Schwarz criterion		8.376085
Log likelihood	-1347.941	Hannan-Quinn criter.		8.334201
Durbin-Watson stat	1.978079

Variables used in OLS regression

D1INV = the difference between the change in inventory this year and the change in inventory in previous calendar years.

D1HDDS = difference between heating degree days in the current week and same week from previous years

D1HDDS2 = the difference between the square of heating degree days …

D1CDDS = = the difference between cooling degree days…

D1CDDS2 = the difference between the square of cooling degree days…

Trend(-1)= Residual term from previous period. (Since the market cannot respond immediately to an imbalance, the imbalance is somewhat autoregressive)

Trend5= Average of residuals from previous periods 2 through 5.

Resources:

¹ For the intrepid reader who wants a more rigorous source, the Peter Kennedy book on Econometrics is among the best.

Copyright © 2010 Matthew Millar
Editorial Archive

Matt has been fascinated with economic, energy, and financial issues since 2003. Leaving his first career as a cellist with the Seattle Symphony in 2005, Mr. Millar enrolled in a graduate program in economics at the University of Oregon. There, he focused his attention on energy issues, international finance, and time-series econometrics, and received his Masters degree in 2009. Concurrent with his pursuit of this degree, Matt started and ran an incubator hedge fund. His professional designations include the FINRA Series 3 commodities and futures license. Currently, he manages money for "Through The i," writes articles on macro economic issues, and teaches.

Contact Information

Matthew Millar | Manager, Through The i | Email