How seasonality adjustments work
A seasonality adjustment lets you tell Smart Bidding, in advance, that conversion rate will rise or fall by a set percentage over a specific short window. During a flash sale you might signal that conversion rate will jump 40 percent for 48 hours. Bidding then leans in immediately rather than waiting to observe the spike and react a beat too late. When the window ends, bidding returns to normal. The adjustment changes how aggressively the algorithm bids during the event, not what it counts as a conversion.
When to use them, and when not to
Seasonality adjustments are designed for brief, high-confidence events of roughly one to seven days: a launch, a flash sale, a promotion with a hard start and end. They are not for routine seasonality like Black Friday demand curves or the summer slump, because Smart Bidding already models recurring patterns from history and a manual adjustment on top can double-count. They are also wrong for long sales, where the algorithm has enough time to learn the new rate on its own. Overusing them fights the model instead of helping it.
Seasonality adjustments vs data exclusions
The two tools point in opposite directions. A seasonality adjustment is forward-looking: you know a spike is coming and you prime bidding for it. A data exclusion is backward-looking: something broke conversion tracking for a period, and you tell Smart Bidding to ignore that stretch so a tracking outage does not poison its learning. Confusing them causes real damage, since excluding a genuine sales period or priming for an event that never happened both teach the algorithm something false.
How iClick uses seasonality adjustments
iClick reserves seasonality adjustments for short, high-confidence events with a hard start and end, and lets Smart Bidding learn ordinary seasonal patterns on its own. The rule is to apply an adjustment only when there is a specific reason to expect a conversion-rate move the algorithm cannot anticipate, such as a launch or flash sale, and to size it from prior comparable events rather than guesswork. Used sparingly they sharpen bidding, and used constantly they just fight the model.
