"Trials and Tribulations of Predicting the EIA Natural Gas Storage Number"
From RBN Energy:
Last Thursday (January 31, 2013) a high frequency trading event about
400 milliseconds before the scheduled EIA Storage data release set off a
frenzy of selling in the NYMEX Natural Gas futures market. Some
analysts believe that the storage report could have been leaked. Others
blamed computer timing issues. These events resurrected ghosts of EIA
storage numbers past and inspired our examination of these developments
in the context of the evolution of high frequency trading in energy
markets.
The EIA Storage Number
Every Thursday at 10:30am Eastern Time, the Energy Information
Administration (EIA) releases the Weekly Natural Gas Storage Report,
definitely the most widely followed piece of fundamental data monitored
by US natural gas traders. The report is seen as a snapshot of the
demand / supply balance in the US natural gas markets. High levels of
injections in the summer (normalized for weather conditions) are an
indication of robust supply and/or lagging demand, and send a bearish
signal to the market. High levels of withdrawals in the winter (again,
adjusted for weather) are likely to produce a spike in natural gas
prices. Thursday morning is also a moment of truth for the entire
cottage industry of storage forecasters whose weekly predictions are
tabulated, distributed, ranked, sometimes praised and/or ridiculed by
the industry newsletters and news reporting organizations [1]. The
quarterly ‘accuracy’ ranking (an equivalent of a school report card) may
break or make the career of an aspiring fundamental analyst or natural
gas trader. The storage number is often misinterpreted by end users.
And according to some observers, the number’s release is increasingly
being used as a window for high frequency traders to “attack” the
market. We’ll explore these issues below.
First of all, this is not a report of storage injections or
withdrawals. The report is an estimate of working gas in storage and
weekly flows in and out of storage are implied by week-to-week changes
of the overall levels. Injections and withdrawals account for most of
the inventory level changes but other factors may intervene from time to
time. Unavoidable noise in the data comes from accounting changes
(reclassification of cushion gas [2] as working gas or vice versa) and
potential mistakes. The weekly report is based on a sample of natural
gas storage facilities. It is a very large sample (relative to the
entire population of natural gas storage facilities) but a sampling
error is unavoidable. Anybody who has been in this business long enough
knows that these factors can intervene when they are least expected and
can inflict a lot of pain on unsuspecting traders and others who rely
excessively on these statistics.
The Enron Days
During my Enron days [Vince was Managing Director for Research and Quantitative Modeling at Enron from 1992 to 2002 – rb],
the group I managed was responsible for forecasting storage levels
(reported at the time by the American Gas Association (AGA). The storage
number was just as important back then, and it was a big factor in the
development of the company’s gas trading strategies. I was called on
the carpet one day by my boss and taken to the woodshed for producing a
prediction diverging widely from the reported storage level for the
week. I had a good explanation: the reported number was likely to be
wrong and the trading floors from Houston to New York City agreed.
This excuse was not good enough. My boss told me that he did not want a
forecast of true storage numbers. My job was to predict AGA mistakes.
Additional excuses did not work either. Yes, we could predict with some
success systematic, i.e. regularly occurring, mistakes. We knew from
experience, I told my boss, that the storage numbers are often incorrect
around long weekends (the 4th of July, Thanksgiving,
Christmas). Not only does the demand diverge from typical levels,
throwing a monkey wrench into quantitative models, but often the
pipeline’s employees in charge of submitting storage numbers are on
vacation. They might submit a guesstimate or repeat the previous week’s
storage data. Eventually, those errors are corrected in the next
submission but the storage numbers are off the mark two weeks in
succession. “Not good enough,” my boss told me. “You have to predict a
mistake each time it happens and do it better.” Oh, those were the days my friend (link to sing along video)....MORE
