I tried to mimic the benchmark algorithm by executing a simple buy and hold strategy of SPY. The same algorithm is used in this thread: https://www.quantopian.com/posts/backtester-change-updated-benchmark
As explained in the post of the link, there can be differences between both algorithms because it is not possible to buy partial shares of SPY and thereby use up all cash and because the benchmark is fully invested at market open, whereas SPY needs to be bought. So, in the following plot I understand that there is a discrepancy at the start of both algorithms. However, why is the difference suddenly increasing on 12/15/2016?