Algorithmic Trading

Goal:

Select a supervised algorithm that can predict stock prices of historical data based on the predictors i.e. statistical indicators - like Volatility, Bollinger Bands etc.

H2OAlgorithmicTrading.py

Accordingly formulate a trading strategy (based on predicted values) in order to generate orders dynamically (on same historical training set for backtesting) and observe the gain / loss in overall portfolio.

tradingSimulator.py , orderGeneraor.py

Once we are satisfied with the results , we are ready to run the same program for 'real-time trades' FInally, we nned to select the combination of Machine learning algorithm and Trading strategy to maximize gain for future orders.

Run the Application:

python H2OAlgorithmicTrading.py python StandardAlgorithmicTrading.py

Assumptions:

h2o python modules installed (see Referenced at the bottom of the page)
write permission on /tmp/ folder and write permission on current working folder
faced problems downloading files from H2O server and converting pd frames from/into H2O frames
will work only in python 2.7
if one wants to point to local / remote h2O server, then comment out h2O.init() and uncomment h2O.init(ip=..,port=...)

Performance of H2OGeneralizedLinearEstimator(family = "gaussian")

Algorithm Metrics:

Dataset: Link : https://drive.google.com/file/d/0ByhSuUifwO07bms4NmFNNjVwbTQ/view Symbol : IBM Training Data: pd.date_range('2008-01-01', '2009-12-31') Testing Data: pd.date_range('2010-01-01', '2010-12-31')

MSE: 0.0011164438007 R^2: -0.0556747669549 Mean Residual Deviance: 0.0011164438007 Null degrees of freedom: 251 Residual degrees of freedom: 249 Null deviance: 0.269232404456 Residual deviance: 0.281343837775 AIC: -989.851898711

Trading Results:

Initial Portfolio Value: 100000 Data Range: 2008-01-01 to 2009-12-31

Sharpe Ratio of Fund: -0.868324247687 Sharpe Ratio of SPY: -0.137143768005

Cumulative Return of Fund: 32.3673 Cumulative Return of SPY: -0.194324631101

Standard Deviation of Fund: 1.98264249049 Standard Deviation of SPY: 0.0219321223021

Average Daily Return of Fund: -0.108449237284 Average Daily Return of SPY: -0.000189476626317

Final Portfolio Value: 333673.0

Performance of standard numpy Linear Regression Classifier - np.linalg.lstsq(dataX, dataY)

The main observation is variation of predictions from actual value considerably high and as a result the particular trading strategy does not work for this standard algorithm very well.

Strong prediction automatically bossts up the Trading as the decision to but / sell is based on a threshold on the difference between actual and predicted values. Definitely we can make this strtagey more robust.

Algorithm Metrics:

In sample results RMSE: 0.0371178709308 corr: 0.499644155508

Out of sample results RMSE: 0.0199253542973 corr: 0.488509112306

Trading Results:

Initial Portfolio Value: 100000 Data Range: 2008-01-01 to 2009-12-31

Sharpe Ratio of Fund: -0.311980176564 Sharpe Ratio of SPY: -0.137143768005

Cumulative Return of Fund: -38.1258 Cumulative Return of SPY: -0.194324631101

Standard Deviation of Fund: 1.12516858177 Standard Deviation of SPY: 0.0219321223021

Average Daily Return of Fund: -0.0221127948848 Average Daily Return of SPY: -0.000189476626317

Final Portfolio Value: -371258.0

Issues:

we shouldn't save intermediate results into files, rather should upload into h2o server and download from it
the plots and images are not getting rendered due to OSX issues so commented out time being
file permission issues with IPython , once we interact with h2O cloud for dat atransfer, this problem will go away.