Skip to content

SelectModelStock

Jump to bottom
Keith McDonnell edited this page Aug 16, 2012 · 5 revisions

 

This time, you should classify a stock based on its own time series. A second time series may also be added; eg compare INTL & QQQ

Dimensions

  • (previous min) percent change open
  • (previous min) percent change high
  • (previous min) percent change low
  • (previous min) percent change close
  • (previous min) percent change volume
  • (previous min) percent change open-low
  • (previous min) percent change open-close
  • (previous min) percent change open-high
  • (previous min) percent change low-close
  • (previous min) percent change low-high
  • (previous min) percent change close-high
  • (current min) percent change open-low
  • (current min) percent change open-close
  • (current min) percent change open-high
  • (current min) percent change low-close
  • (current min) percent change low-high
  • (current min) percent change close-high

Arguments

  • data set
  • k range
  • primary stock
  • secondary stock (optional)
  • iterations
  • number of days

eg select_model_stock.py -d train_2012_12...h5 -k 5,10 -i 10 -d 20 -p INTL -s QQQ

Note:

  • primary stock: INTL
  • secondary stock (optional), eg QQQ
  • number of days: select N days randomly from data set
  • iterations: find average after N random trials

Calculate weights

Using a simulated annealing cost function, calculate the scaling weights for all dimensions.

See Collective Intelligence, pg. 181 - 182 for details.

Weights should range from 0 - 1.0.

Output

eg select_model.py -d train_2012_12...h5 -k 5,10 -i 10 -d 20

data set: train_2012_12...h5
k range: 5,10
iterations: 10
days: 20
scale: QQQ %open 1.0, INTL %open-low 0.9, ... INTL %volume 0.
std dev error: 3.4%
avg error: 12.5%

Reading notes

Please apply the traing data generated in PrepareData to the kNN algorithm outlined in Programming Collective Intelligence, ch. 8 .

I've uploaded the source code for the book, including the kNN code you'll need to work with.

You might find scikit-learn a useful library. You can use the book examples, a standard library or roll your own.

In this case we'll be classifying instead of predicting a value.

Clone this wiki locally