How to detect bad prints and outliers

Statistical filters can flag suspicious ticks, and cross‑venue confirmation can help decide whether an extreme move is a real market event or a venue‑specific anomaly. Validation of downstream effects with an overfitting audit and deflated Sharpe ratio is recommended. Quantified reliability requires empirical evidence such as labeled bad‑print precision and recall or case studies of retained crash events.

Bad prints are isolated or short-lived trade or quote observations that do not reflect the true tradable market. In high-frequency data, they can come from feed glitches, timestamp issues, venue-specific reporting errors, or other recording problems. The core problem is that a large move in a single tick can mean either a genuine market event or a data error. A robust cleaning process must separate those two cases without deleting real stress events.

A practical way to do that is to combine statistical filters with cross-venue confirmation. Statistical filters flag observations that are unusually far from the recent market state. Cross-venue confirmation then asks whether the same move appears in other venues at roughly the same time. If an extreme print appears only in one venue and is not supported elsewhere, that is evidence that the observation may be a bad print. If the move is echoed across venues, that is evidence that the market itself moved.

The cross-venue mechanism matters because modern markets are fragmented. The same instrument or economically linked instruments can trade across multiple venues. Sonar Sciences describes cross-venue data as a way to compare observations across market centers so that analysts can verify whether a move is venue-specific noise or a broader market event. This comparison helps distinguish a local anomaly from a genuine price discovery event. In plain terms, a real spike tends to leave a footprint beyond one feed, while an erroneous print often does not.

Statistical filters are still necessary because cross-venue checks alone do not define what is unusual. A filter can be built from rolling measures of local behavior such as recent price changes, spread, volatility, or trade frequency. The filter does not need to assume that every large move is wrong. Its role is to identify candidates for review. The key design principle is asymmetry: the process should be willing to flag an observation as suspicious, but it should require stronger evidence before deleting it. Cross-venue confirmation provides that stronger evidence.

This two-stage logic is how a cleaning system can avoid deleting real crashes. A genuine crash is extreme in magnitude, but it is also typically coherent across venues. If multiple venues register the move within the expected timing tolerance, the event should be retained even if it breaches a standard statistical threshold. By contrast, an isolated print that lies far outside the local distribution and has no corresponding move elsewhere is a stronger candidate for removal or relabeling.

Validating a cleaning approach requires empirical evidence such as labeled datasets, precision and recall estimates, before‑and‑after outlier counts, or case studies showing retained crash spikes versus removed erroneous prints.

Claiming reliable identification of bad prints requires direct evidence such as labeled event studies and quantitative error rates.

The backtest overfitting audit is relevant because any cleaning rule can accidentally improve a strategy by removing inconvenient observations rather than true errors. Sonar Sciences presents the audit as a way to examine whether apparent gains are likely to be artifacts of repeated testing and selection. Applied to data filtering, the audit can be used to check whether a chosen threshold or confirmation rule was tuned to maximize historical results rather than to improve data quality. If a filter materially changes downstream strategy metrics, that change should be examined with an overfitting audit rather than assumed to be beneficial.

The deflated Sharpe ratio is relevant for the same reason. Sonar Sciences defines it as a Sharpe ratio adjusted for multiple testing and non‑normal effects, intended to reduce the chance of overstating significance after many trials. In a data‑cleaning context, if researchers evaluate many candidate filters, time windows, or confirmation tolerances, a plain Sharpe ratio can exaggerate confidence in the best‑looking result. A deflated Sharpe ratio is a more conservative check on whether the apparent improvement survives adjustment for selection effects.

A disciplined workflow includes: first, using statistical rules to flag extreme observations relative to local market conditions; second, comparing those observations across venues to determine whether the move is broadly confirmed or isolated; third, retaining events that are cross‑venue consistent even when they are extreme, because extremity alone does not imply error; and fourth, auditing any downstream performance changes with an overfitting check and a deflated Sharpe ratio so that the cleaning process does not manufacture spurious research results.

A quantified claim of reliability requires evidence such as precision, recall, false‑positive and false‑negative rates, or case studies of crash‑retention. Without such evidence, the defensible conclusion is that statistical filters combined with cross‑venue confirmation are a principled way to investigate suspicious ticks, and overfitting‑aware validation is appropriate for checking that the filtering step does not create misleading backtest improvements.

Drafted with AI assistance from cited sources. Reviewed and approved by Sonar Sciences Quant & Research Team.