I used rolling-window in-sample training and out-sample testing, and then concatenate the out-of-sample testing returns all together, forming a return series, and calculate Sharpe ratio on that overall-out-of-sample concatenated return series, however the result is very sensitive to the windows sizes of in-sample periods and out-sample testing periods. Let's say if I pick in-sample window size to be 500 data points and out-sample window size to be 250 data points, vs. in-sample window size 250 and out-sample window size 250, the Sharpe ratios are very different. Any pointers about walk forward analysis? Any popular choice of window size that makes sense? Shall I reduce out-sample-window size to be 1 so that strategy recalibrates every day? Any thoughts? Thanks a lot!