Why two EMAs disagree: reconcile warm-up history and seeds

Distinguish formula readiness from numerical agreement

An exponential average follows Eₜ = α × Pₜ + (1 − α) × Eₜ₋₁. To replay it you need α, the price source, ordered observations and an initial state. Implementations may seed with a simple average of an initial window or use another declared convention. Loading more historical observations can therefore change the value on the same latest candle.

A library’s ready flag answers that library’s initialization requirement. It does not prove two histories agree within a tolerance you care about. First compare instrument, venue, adjustments, candle timestamps, input field and missing-bar handling. Only after aligning those inputs should you attribute a difference to the seed.

Calculate how much starting-state difference remains

For two EMAs with the same fixed α and identical subsequent inputs, subtract their recurrences. The input cancels, leaving Dₜ = (1 − α) × Dₜ₋₁. After k updates, Dₖ = D₀ × (1 − α)^k. This identity measures the remaining seed difference, not either series’ distance from a true price or an optimal forecast.

Choose a comparison tolerance before looking at a trading outcome. A tolerance in absolute price units has different implications across instruments; a relative or tick-based tolerance also needs a declared reference. If inputs later diverge, the seed-only equation no longer accounts for the entire discrepancy. Rounding each intermediate update can introduce another difference.

A nine-period EMA needs 21 identical updates for this chosen tolerance

Use α = 2 / (9 + 1) = 0.2. Two valid prior histories leave current EMA states A = 100 and B = 90. From that point, both receive the same hypothetical close of 100 on every update. A stays at 100; B follows 100 − 10 × 0.8^k. These are deliberately assigned starting states, not an assertion that a particular platform seeds this way.

After 20 updates the difference is approximately 0.115292, which exceeds a predeclared 0.1-price-unit tolerance. After 21 it is approximately 0.092234, so 21 is the first sufficient integer count for this example. A nine-period name does not mean all earlier influence disappears after nine bars. More history is a reproducibility choice, not evidence of a better trading edge.

More historical bars cannot fix a different price series

One chart using mark-price closes and another using last-trade closes need not converge to each other merely because both load a long history. The cancellation argument requires identical inputs at every compared update. Check that prerequisite before endlessly increasing a warm-up window.