We've looked at a lot of possible avenues for finding and fixing issues in the simulation and historical market data playback framework. If we are still observing differences in trading strategy performance in live markets as compared to simulations, then another possible solution to explore would be to adjust the expected performance results as obtained from simulations to account for the backtester bias.

As we discussed before, the backtester bias can be optimistic or pessimistic in nature and can be a constant bias or a bias that varies by trading strategy type, by strategy parameters, or by market conditions. If the bias can be isolated to be constant for a specific strategy type and strategy parameters, then it is able to collect simulation dislocation results from live trading results and organize them per strategy and per strategy parameter set. These expected dislocation values can then be used with the simulated results to estimate true live trading results. For example, if an algorithmic trading strategy with specific parameters always performs 20% worse in live trading as compared to simulation results because of simulation dislocations, we can account for that, reduce its simulated results by 20%, and re-evaluate it. We can take this estimation methodology one step further and try to model the magnitude of backtester optimism/pessimism as a function of traded volume and market conditions, such as how busy the market is or how much the prices changed.

In this manner, it is possible to build a system that takes simulated results for trading strategies and then takes live trading results for the same strategies and tries to quantify the simulation dislocations and provide estimates of true expected live trading performance. These methods of adjusting expected live trading performance are not ideal; they require feedback from running trading strategies in live trading, which might cause losses and, at the end of the day, is just an estimation. Ideally, we want a backtester capable of providing accurate simulation results, but since that is an extremely difficult and sometimes impossible task, this estimation method is a good middle ground for dealing with simulation dislocations and continuing to build up and manage an algorithmic trading business.