In systematic investing, the ability to distinguish an unusual observation from a broken one is not a data-cleaning detail. It is part of the research problem.
Quantitative strategies begin with data.
That makes data quality sound like an infrastructure problem.
Clean the dataset.
Fix missing values.
Remove obvious errors.
Move on to the interesting research.
In reality, the boundary is much less clean.
Sometimes the unusual observation is the thing the researcher should remove.
Sometimes it is exactly the thing they should investigate.
Distinguishing between the two can be commercially important.
The dangerous error is the plausible one
A value that is obviously impossible is relatively easy to catch.
A price of zero where zero cannot exist.
A timestamp decades out of place.
A malformed identifier.
Those problems are annoying.
The harder problems are plausible.
A sudden jump that could represent an event.
A strange pattern created by a vendor methodology change.
A missing observation that looks like information.
A historical value that was revised later.
A feature that appears predictive because of how the dataset was assembled.
These problems can survive basic cleaning.
Worse, they can create apparently strong research.
Data problems can create false confidence
A backtest does not know whether the dataset represents the information that would actually have been available at the time.
It only knows what it has been given.
That creates familiar risks.
Look-ahead bias.
Survivorship bias.
Restated data.
Timestamp inconsistencies.
Selection effects.
Vendor changes.
Corporate actions.
Incorrect joins.
Small implementation details can turn into very convincing performance.
That is why serious quantitative research requires understanding where data came from, how it changed and what it actually represented at each point in time.
Alternative data makes the problem harder
The challenge becomes even more obvious with large, messy real-world datasets.
Alternative data can include transactions, web activity, satellite imagery, shipping information, app usage, text, geolocation and countless other sources.
The attraction is clear.
The data may reveal something that traditional market datasets do not.
But the further the information moves from a clean exchange feed, the more interpretation becomes part of the research.
What produced the data?
Why is coverage changing?
Is the sample representative?
Has the vendor changed collection methodology?
Could the apparent signal simply be a change in measurement?
Understanding the dataset becomes inseparable from understanding the model built on top of it.
This changes who creates value
It is one reason the line between data engineering, quantitative development and quantitative research can become blurred.
A researcher may discover the statistical relationship.
A developer may discover that the underlying value was corrupted.
A data specialist may realise the historical dataset was reconstructed using information unavailable in real time.
Any of those observations can determine whether the strategy is real.
The most important contribution may therefore happen before the model is built.
Data scepticism is a research skill
Good researchers tend to develop a particular suspicion towards clean-looking results.
Where did this number come from?
When would we have known it?
What changed in the collection process?
What happens if we remove this period?
Why does coverage suddenly improve?
Could the model be learning the vendor rather than the market?
Those questions can feel less glamorous than discussing model architecture.
They often matter more.
The lesson for hiring
When assessing candidates who have worked heavily with data, asking which technologies they used only goes so far.
A more revealing discussion is:
Tell me about a dataset you initially misunderstood.
What was wrong with it?
How did you discover the problem?
How did it affect the research?
What checks did you build afterwards?
That moves the conversation from tools to judgement.
In quantitative investing, that distinction matters.
Because bad data does not always look broken.
Sometimes it looks like the best backtest in the room.
Key takeaway
Data quality is part of the investment process. The ability to understand how information was produced, recognise misleading observations and distinguish genuine signal from measurement error is a core quantitative research capability.