Filtering Outliers in Crypto Price Feeds
Identify suspicious price observations with explicit thresholds and market context, while preserving raw values and avoiding hidden changes to historical series.
Overview
When translating these concepts into an integration, keep the scope described by the wallet history api with the result. The examples in wallet transaction history and smart contract event logs show how the same distinction affects neighbouring data questions.
An extreme price can be a genuine market event, a unit error or a poor observation. An outlier filter should record its reasoning rather than silently rewriting history toward a preferred value.
Check structure before statistics
Validate asset identity, pair orientation, decimal scales and observation time first. A thousandfold jump may be a unit conversion error rather than a market move. Statistical rules then need a declared comparison window and minimum evidence count. Sparse markets require different caution from frequently traded ones.
Worked example: an inverted quote
A feed normally reports quote units per base. One input arrives in the inverse orientation and appears as an extreme low. A median-based filter might hide it, but a schema-aware check identifies the actual integration error. Preserve the rejected observation and reason so the parser can be corrected.
Publish raw and selected states
Keep accepted, flagged and rejected statuses with rule versions. Test legitimate sharp moves, a single-trade market and correlated bad inputs. A filtered price is a derived result, not a replacement for the original evidence. Monitor how much data the filter removes; persistent exclusions can indicate a broken input contract rather than improved market quality.
A practical acceptance test
Supply an inverted pair, a decimal-scale error and a genuine large market move. Structural validation should identify the first two before a statistical rule runs. The genuine move may be flagged but must retain its original value and reason. Compare raw and selected series so a filter change cannot silently rewrite the evidence being evaluated.
Questions about the topic
Frequently asked questions
Is every extreme price wrong?
No. Real markets can move sharply.
Should rejected observations be deleted?
Retain them under an appropriate retention policy for diagnosis and reproducibility.
Standards & technical documentation