John Jarvis, Claudia Johnson & Liana Vetter
May 6, 2004
Description of Problem
• Quest’s current gas marketing– Oneok is sole purchaser
• 85% guaranteed monthly• The remainder sold daily
– Pipeline serves as middleman
• Goal of the project– Analyze the market trends and forecasting
accuracy of Quest– Determine what percentage is optimal to
guarantee on contract– Create optimization model Quest can use
monthly
Variables Considered
• Two different sale points– R&H: large and unstable
– Housel: small and unstable
• Historical data – Forecasted daily production by sale point (2004)
– Actual daily production by sale point (2004)
– Daily NYMEX prices (2002-2004)
• Limits to set– Maximum days and amount in debt
– Bounds on percentage to guarantee
Limits to Set
• Maximum days and amount in debt– Set limit of 2 days in debt based on 2004 data
– Set limit of 10% of production in debt
– Conservative limits to minimize risk in case of unexpected changes in production
• Bounds on percentage to guarantee– Set upper limit as 95%, highest Quest has used
– Set lower limit as 30% to protect against sharp decrease in production
Market 2002-2004Market Variability
0
1
2
-0.38 -0.13 -0.03 0.02 0.07 0.16 0.26
Percent Market Change
Fre
qu
en
cy
Falling
Rising
Probability Average percent changeFalling 0.44 -0.055751Rising 0.56 0.096644
R&H Production 2004
RH Probability Average relative errorOver 0.51 1.073852Equal 0.27 1Under 0.22 0.900982
R&H Production Variability
0
2
4
6
8
10
0.57 0.83 0.9 0.95 0.98 1.01 1.04 1.07 1.1 1.14 1.17
Relative Error on Forecast
Fre
qu
en
cy
Under
Equal
Over
Non-Stochastic Model• In 5-day test case, user provides data:
• Model returns output:
• Revenue: $1,002
Day 1 2 3 4 5 Production 130 90 80 110 100 Daily price $1.50 $3.00 $1.75 $2.10 $1.00 Monthly price $2.00
Day 1 2 3 4 5 Monthly guarantee 90 90 90 90 90 Daily sales 40 0 0 20 0 Total sales 130 90 90 110 90 Debt 0 0 10 10 0
Stochastic Model• Benefits of stochastic modeling
– Incorporates uncertainty using probabilities of different scenarios– Calculates expected revenue based on market forecasts– Approximates actual production from forecast given
• Example case– User provides data:
– Model returns output:
– Expected revenue: $1,039
Probability market rises 0.8 Probability market falls 0.2 Price at beginning of month 2 Forecasted daily production 100
Monthly guarantee 30 Approximate daily sales 71 Expected Production 102
Stochastic Model with Regret• Regret – difference between optimal revenue and actual
revenue• Benefits of regret
– Solution does well in rising and falling market– Less sensitive to predicted probabilities
• Example case– User provided data:
– Model returns output:
– Expected revenue: $1,037
Probability market rises 0.8 Probability market falls 0.2 Price at beginning of month 2 Forecasted daily production 100
Monthly guarantee 39 Approximate daily sales 63 Expected Production 102
Sensitivity Analysis• Optimal monthly guarantee varies little when expected
production data changes
• Model is more sensitive to changes in market data
R&H Market Sensivility
0
500
1000
1500
2000
2500
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Probability of Rising Market
Op
tim
al
Mo
nth
ly
Gu
ara
nte
e
Problems and Limitations
• Problems encountered– Limited historical data
– Multiple daily gas prices (strip price used)
– Large variability of the gas market
– Difference in production records from meter inconsistency
• Limitations of the solution – Dependant on the market which is unpredictable
– Stochastic variables are based on limited data
Analysis and Recommendation
• 50-55% should be guaranteed monthly if no market predictions added from Quest
• Consequences of guaranteeing 50-55%– $18,000 additional revenue from January – March
2004 for R&H– $2,400 additional revenue from January – March
2004 for Housel
• Regret model yields less additional profit than stochastic model but provides more consistency between months
Interface
• Questions asked by interface– Probability the market will rise– Sale point– Month to forecast, days in month– Expected initial NYMEX price– Forecasted daily production– Expected beginning debt
• Results of interface– Creates data file for AMPL– Data file can be run with regret model to resolve
each month