That sounds about right, and even 4 years of data is kinda small. Years ago I made and SPY trend following algo and optimized it in one year of data, thirteen years, everything in between, and let it run into the future to check the results. The one year optimized version made no money. The thirteen year one (all the data I could download at the time) did the best. Basically the equity curve kept going up at the same rate as in backtesting, but I didn't test it with real money. As an anecdote, I never really figured out discretionary trading until about 12 years. Now it's just psychology holding me back, but isn't this always the case?