>The funnest part of writing a bot is using all the historical data to observe how the bot would have done during that period. I ended up checking thousands of permutations of parameters by brute force to determine the best ones, but in the end transaction fees killed me.

That's a sure way to find a bad strategy. You end up doing the equivalent of overfitting. If you look at enough variables, the chance of there not being spurious correlations is basically zero, so you're guaranteed to land on sets of parameters that look good with historical data but fail going forward.