Predicting Commodity Prices Using Artificial Neural Networks

1
Predicting Commodity Prices Using Artificial
Neural Networks
Andy Korth
2
Outline

• Description of the problem
• Acquiring the dataset
• Understanding the data
• Design of neural network
• Challenges and revisions
• Interpreting the results
• Conclusions

3
Economic Predictions with Neural Networks

• Moderately successful in many cases, such as
  weekly US Treasury Bond purchases 1
• 67 accuracy on buy predictions 1
• Generally considered powerful tools
• Success is a function of amount of historical
  data in training set, ... and how data is
  preprocessed.
• Models have a finite time of accuracy, before
  needing to be retrained

1 Cheng, Wei, et all. (1996). Forecasting the
30-year U.S. Treasury Bond with a System of
Neural Networks
4
Economic Systems

• The World of Warcraft (WoW) economy mirrors real
  economies in many ways 2,3,4
• WoW primarily uses an auction system.
• Bid or Buy it now prices available

2 Economic Integration Strategies for Virtual
World Operators, Lehdonvirta, 2005 3 Bartle,
Richard A. (2003). Designing Virtual Worlds.
Indianapolis New Riders. 4 Castronova, Edward
(2002). On Virtual Economies. CESifo Working
Paper Series No. 752.
5
(No Transcript)
6
The Dataset

• Data gathered February - April 2006
• 69 scans, totaling 201,309 data points
• I wrote Java software to parse the data files

7
Pre-processing and Outliers

• Outliers can be a significant problem
• Remove data outside of 2 standard deviations
• After removing outliers, calculate averages
• Linearly normalize data
• Represent certain concepts as true or false
• isWeekend
• isEvening

8
Initial Design

• Inputs
• Buyout Price
• Bid Price
• Number Available
• Date
• Number with bids
• Minimum price

• Output
• Expected Buyout Price Change
• Expected Bid Price Change
• Expected Volume

9
Limitations of ANNs

• Cannot represent disjoint (non linearly
  separable) information on a single input
• Such as day of the week

I0 T1 I1 T0 I2 T1
Cannot be learned on a neural network! (in this
form)
Cite nn book
10
Solution Multiple Inputs

• Make boolean inputs.
• To represent every day in the week, you will need
  seven inputs
• Recall our identity assignment
• There were 8 inputs, instead of one input from 0
  to 128

11
Analyzing data
I0.5553250345781466 0.22853157999601514 0.9 1 T
0 I0.6307053941908713 0.23291492329149233 0.85
0 T 0 I0.5235131396957123 0.21956565052799362
0.85 0 T 0 I0.553941908713693
0.22952779438135087 0.9 1 T 1 I0.82641770401106
5 0.343893205817892 0.9 1 T 0 I0.5
0.23709902370990238 0.8 0 T 0 I0.52835408022130
01 0.2532376967523411 0.9 1 T 0 I0.625172890733
0567 0.23829448097230524 0.85 0 T
0 I0.6189488243430152 0.25303845387527396 0.85
0 T 0 I0.6127247579529738 0.2502490535963339
0.8 0 T 1

• Note the poor distribution of numbers in the
  buyout and bid (first two) columns
• Around 40 of training examples incorrect

12
Normalizing Data

• Better results can be found with inputs well
  distributed between 0 and 1.
• Remove outliers (outside of 2 std. Dev)
• Find the max and min values in each set
• Scale the largest value to 1 and the smallest to
  zero

13
Adjusting Inputs and Outputs

• Inputs
• Avg Bid Price
• Avg Buyout Price
• Quantity Avail
• Min Bid Price
• Min Buyout Price
• isWeekend
• isEvening etc..

• Output
• Price above average or not

14
Results of Improved Dataset

• Better Distribution!
• Low error on major commodities
• 11 incorrect on Runecloth, 20,000 cycles
• 8 incorrect on Dreamfoil, 25,000 cycles
• 9 incorrect on Plaguebloom, 250,000 cycles
• Less accurate on less heavily traded valuables
• 25 incorrect on Arcane Crystal, 25,000 cycle

Learning rate 0.1, Momentum 0.05
15
Conclusions and Future Work

• Neural Networks are highly effective in
  predicting market prices in WoW
• Already successful in real life tests
• 80 gold profit in one day (a few USD)
• Room for improvement
• Better output variables
• Better statistical analysis (use medians)

16
(No Transcript)