# What's the best way to measure accuracy/significance of a Stock Prediction Model?

I'm looking for insights on how to test the accuracy of a model I built to predict next days stock price - Open, High, Low and Close next day prices.

Is there a unique indicator of accuracy or standard approach that should be used to represent the accuracy of this system ?

This is a predictive model for prices, not a trading system.

Input to the model is the historic time series - end of day stock prices-Open, High, Low, Closing.

I've run some initial stats by calculating the ratio of predicted prices to the actual end of day prices i.e if PC represents predicted close price for next day I am simply taking the ratio of PC / C where C is the actual close price for the predicted day and representing it as a %.

Have researched some forecasting system models online and understand that the common approach is to use Mean Absolute Percent Error, Mean Absolute Deviation, R-Squared, Correlation, Cumulative Forecast Error. As an example my prediction model is giving me values for Correlation and R-Squared 0.995 0.9924 0.9912 0.9885 & 0.9905 0.983 0.981 0.976 respectively for the O H L C predictions..would these be considered good values ?

Any suggestions on what would be the best approach?

• As written, some of these issues might be a better fit to a different forum, perhaps one related to investment (such as questions about what's 'standard'). Additionally, you shouldn't expect there to be a single best measure unless you specify some criterion by which to judge 'best'. – Glen_b Nov 13 '13 at 6:35
• This may sound flippant, but it's a serious point: if it's just a "predictive model," then who cares? The determination of how to assess a model comes down to how it will be used. If not for trading, then for what? Advancing our understanding of economic systems? Providing strategic analysis for a corporation? In the former case you will be concerned about the reliability and generalizability of estimates of the parameters in the model. In the latter case you might be most concerned about the accuracy of prediction far into the future. What is the objective of your model? – whuber Nov 13 '13 at 15:55