How do you handle backtesting models that seem to perform well on historical data but fail live?

Forward testing often reveals hidden variables like spread, execution speed, and emotional friction. How do you refine your strategy when forward-testing results diverge from backtesting?