How can I tell if my strategy is overfit?

Five practical smells: (1) performance collapses when you split the data and test only on the held-out part; (2) tiny parameter changes cause big performance swings — a real edge is a plateau, not a spike; (3) the strategy has many conditions or 'magic' constants that exist to dodge specific historical losses; (4) it works on one coin and nowhere comparable; (5) you tested many variants and kept the best without correcting for the search.

The systematic versions of these checks are exactly what a robustness scorecard runs: out-of-sample splits, parameter-sensitivity sweeps, walk-forward segments and a deflated Sharpe. If a strategy passes all of them it can still fail live — but the failure odds drop enormously.