Fit the change in the spread on its lagged level: delta x(t) = lambda x(t-1) + e(t). The half-life is -ln(2)/lambda. A lambda of -0.10 gives a half-life of 6.9 periods; a lambda of -0.02 gives 34.7 periods.
Use it to set holding limits and to sanity-check a design. If the measured half-life is 30 days and your exit rule gives the trade five days, you are mostly harvesting noise and paying slippage for it. If the half-life is longer than your financing window or your patience, the trade is not viable regardless of how good the statistics look.
Half-life estimates are unstable and regime-dependent. Compute it on a rolling basis and treat a sudden lengthening as a warning that the relationship is breaking rather than as a reason to hold longer.
Related: mean-reversion, unit-root, hurst-exponent, cointegration