ANN Approach to Revenue or Profit Estimation

1
ANN Approach to Revenue or Profit Estimation

• University of Wisconsin-Madison

2
Initial Design Considerations

• Initially, I attempted to create a ANN that would
  predict future stock prices based upon
• Previous Stock Growth Percentages
• Financial Statistics of a firm

• I quickly found that in todays market, stock
  prices may have little in common with a firms
  financial outlook and health.
• Examples
• Ebay,Amazon,Biotech Industry, etc.

3
Initial Design Considerationsltcontdgt

• The next move was to narrow the focus of my
  project to a growth indicator that could be
  considered a derivative of financial health.
• I settled on Revenue growth
• Note This network could easily be changed to
  support profit growth prediction as well other
  financial components. The only restriction is the
  need for enough supporting data.

4
Initial Design Considerationsltcontdgt

• The final consideration was upon the data itself.
  
• First, I could only find inexpensive (read FREE)
  historical financial data for the last three
  year.
• Second, I knew that the training of the network
  would require data for only one industry at a
  time. Basically, each industry operates
  efficiently with slightly different expectations
  on capital asset liability, expected ratios of
  short-term to long-term debt, and various other
  indications. To include a firm from a separate
  industry would be to train or test that firm
  unfairly.

5
Data Gathering

• I needed to gather standardized and extensive
  financial data on each firm for this project. For
  this I ended up using two sources.
• http//biz.yahoo.com/research/indgrp/ provided me
  with stock grouping by industry
• http//www.money.net provided me with the
  information resource of historical financial
  information. Unfortunately their data only went
  back 3 years. A more cumulative training would
  have been interesting.

6
Network Design

• I choose to use a MLP for this project because of
  my familiarity with it as well as its
  flexibility.
• Notes on the specific design of the MLP
• Sigmoid Activation Function
• 25 inputs for year against year training
• 50 inputs for comparative training (includes
  data from both firms)
• 1 output for all designs

7
Input Choice

• I choose my input directly from the financial
  information offered on money.net
• The following data was taken from each firms
  Income Statement, Balance Statement, Cash Flow
  Statement.

• Revenue
• Operating Expenses
• Operating Income
• Income Before Taxes
• Income Taxes
• Pri/Bas EPS Ex. Xord
• Dilutd EPS Ex. Xord
• Primary/Basic Av. Share
• Total Current Assets
• Total Assets

• Total Current Liabilities
• Total Liabilities
• Total Equity
• Net Income
• Depreciation Amort.
• Total Operating Cash Flow
• Total Investing Cash Flow
• Total Financing Cash Flow
• Net Change in Cash
• Receivables

• Accounts Payable
• Common Dividends/Shr.
• Outstanding Shares

8
Output Choice

• For the year vs. year testing and training, each
  firms revenue is compared with that firs
  revenue for the following year. The percentage
  growth is calculated and if it is higher than a
  threshold X, in this case 10 we set that firms
  target output to be 1.
• In comparative testing, we must take the ratio of
  one firms revenue growth versus another firms
  revenue growth. If firms A (located in the outer
  loop) has a higher of growth the target output
  will be 1, otherwise if firm B has the higher
  than the target output will be 0.

9
Results

• While training with both the 1997 and the 1998
  data I was able to obtain an average revenue
  prediction rate of 61.3 over about a hundred
  trials.
• The max of my testing was as high as 70 and the
  low was about 52 .
• Training with the 1997 data and testing on 1998,
  my network obtained an average of 63.4 with a
  range of 70.2 to 54 .
• Training with the 1998 data and testing on 1997,
  the network scored an average of 60.5 with a
  range of 65 to 52.

10
Conclusions New Insights

• The results that I obtained were competitive in
  the sense that the pointed to the ability of a
  neural net to pull viable conclusions for the
  data. While these conclusions are not of the
  investment caliber, I believe the fault of this
  resides with the data itself. Adding the element
  of anticipation or expectation into the data by
  adding a whisper number or expected increase data
  column should help to push the data in the right
  direction.
• The reason that I did not initially propose this
  is because retrieving the PAST data on whisper
  number on a large number of stocks wouldve been
  nearly impossible. Second whisper data tends to
  be extremist. Often it overshoots the true growth
  or decline of a firm. However with the bulk of
  the data predicting conservatively, a whisper
  data field might be a good tweak. (as an aside,
  many of the mis-predictions were clustered around
  the target boundary line. This points to the
  validity of the data for not predicting wild
  results and further suggests a whisper data point
  might help.

11
Conclusions New Insights

• The second change that I would like to make to
  the network is the target space. The real point
  at which a firm becomes investment grade is when
  its growth is expected at gt than 15. Also there
  is a fair amount of clumping at 10 growth since
  that is the industry norm.
• Splitting the Output space into perhaps three
  outcomes of
• Firm X lt5 (sell)
• 5lt Firm X lt15 (hold)
• Firm X gt 15 (buy)