I'm using time series data (continuous features) for binary time series prediction (one step ahead, up-turn and down-tern of output of t+1 comparing to t - t is time). This is number of samples in every year:
Total Up Down
_____ ___ ____
2009 234 135 99
2010 243 153 90
2011 241 132 109
2012 240 133 107
2013 240 155 85
2014 241 110 131
2015 243 126 117
2016 29 24 5
All data 1711 968 743
You can see in above table that we are dealing with unbalanced data-set. After I trained SVM or Neural network models the sensitivity of model is high (70~85%) but the specificity is very low (40~57%). I'm using sliding validation LINK. What is your proposed sampling method for improving specificity (down-trend accuracy) in this case?
