Blog
Blog
Visual temporarily unavailable
This section's visual is under review and a replacement is being prepared.
Bad data does not announce itself. It produces clean charts, plausible backtests, and readings that look precise until they meet the live market.
Data quality is part of market-analysis integrity because every measurement inherits the weaknesses of the feed that produced it. If the feed drops orderbook updates, your depth is wrong. If timestamps drift, your cross-exchange comparison is wrong. If volume includes wash trading, your volume filter is contaminated. The analysis may be logically sound and still operate on a false version of the market.
Most traders audit their indicators. Fewer audit the data beneath them. That is backwards.
Orderbook measurements require a correct local view of the book. That sounds simple until the feed disconnects.
Orderbook updates arrive as a sequence, not a finished picture. If updates go missing or arrive late, the book can show liquidity that no longer exists. A canceled bid can still look available. A depth reading can look precise while describing a market state that has already changed.
The visible failure is rare. The silent failure is common: a reading computes on stale or corrupted state and still returns a number. OFI, depth imbalance, and price impact estimates all look normal because the code ran. The issue is that the input was no longer the market.
The practical question is not whether a system can always return a value. The practical question is whether it can tell you when the reading should no longer be trusted.
Many market-pressure reads depend on order. In fragmented markets, order is not obvious.
If one exchange reports a trade at 12:00:00.100 and another reports a trade at 12:00:00.120, it is tempting to say the first venue led. That may be false. Clocks differ. Feed delays differ. At short horizons, a few milliseconds can invert the story.
This matters for cross-venue analysis. A reading that appears to lead price on another exchange may be a timestamp artifact. A backtest that finds clean lead-lag behavior may be measuring feed delay rather than market structure. The result is especially dangerous because it looks like durable information.
The safer framing is pessimistic: unless timestamp handling is explicit, cross-exchange timing claims are suspect. Exchange fragmentation makes this unavoidable because the same asset trades across independent systems with independent clocks.
Volume looks objective. In digital markets, it is often dirtier than traders want to admit.
A volume total does not tell you whether flow was directional, whether trades represented real inventory transfer, or whether the activity came from participants taking risk. A strategy that treats raw volume as confirmation assumes volume is informative. That assumption fails when volume is inflated, self-referential, or generated without directional intent.
Even when volume is genuine, OHLCV volume loses sequencing and direction. A one-minute candle with heavy volume could contain aggressive buying, aggressive selling, two-sided churn, or absorption by a passive participant. The number alone does not answer the market-structure question.
This is why VPIN and Order Flow Imbalance are useful concepts. They are attempts to recover something more specific than "a lot traded." They ask whether flow was imbalanced, toxic, directional, or likely to change liquidity provider behavior.
Free data usually fails in places that do not show up on the first chart.
The API returns rows. The rows have timestamps. The plot looks continuous. That is not a quality audit. The harder questions are whether gaps were smoothed away, outages were hidden, symbols were mapped consistently, and stale periods were marked as unreliable.
The cost of free data is not only price. It is the false confidence that arrives when defects are discovered after they have already shaped the reading. A backtest built on a quietly patched historical feed can show fills through a period where the live exchange was unreachable. A volume model trained on reconstructed candles can learn relationships that did not exist in tradable time.
Free crypto data is useful for exploration. It is a weak foundation for execution-sensitive research unless the user validates it with the same paranoia they apply to strategy logic.
The calculation is rarely the dangerous part. The dangerous part is trusting a clean-looking number after the market state underneath it has already drifted.
A feature can look precise in the response and still describe the wrong market. Feeds break. Timing drifts. Liquidity disappears during stress. The trader sees one confident reading and never sees the weakness that shaped it.
That distinction matters. A bad calculation can be debugged. A good calculation computed from bad data lies with confidence.
The first audit is basic.
Does the feed preserve gaps or smooth them away? Are timestamps treated consistently enough for the claim being made? Can the system mark a feature as unavailable instead of returning a stale value? Are exchange outages visible in the dataset? Can the same reading be reproduced from time-appropriate data?
These questions are not bureaucracy. They define whether the backtest and the live measurement are describing the same thing. If historical data was corrected after the fact but the live reader never had that correction in the moment, the backtest has an advantage the trader never had.
The honest standard is simple: a reading should be defensible from the data available at the time. If a value depends on future cleanup, future candles, or reconstructed gaps, it is not the same measurement in live analysis.
Data quality is not an internal detail. It is part of market interpretation.
A trader can interpret uncertainty around a noisy reading if they know it is noisy. They cannot interpret uncertainty around a corrupted feed that looks clean. Missing updates, stale timestamps, dirty volume, and reconstructed history turn technical precision into false precision.
Bad data does not just create bad research. It creates confidence in market stories that never existed in the real market.
DepthSignal is useful only if the underlying market state is trustworthy enough to interpret. It is a market-data platform, not a promise engine, and it does not provide financial advice, trading signals, or investment recommendations.
Because the reading is only as trustworthy as the feed behind it. Missing deltas, stale timestamps, smoothed gaps, and reconstructed history all change the market state the model thinks it is reading.
False cleanliness. A corrupted feed that looks precise is more dangerous than a visibly noisy one because it creates confidence in readings that were never available in live market conditions.
Whether gaps are visible, whether order books rebuild correctly, whether stale values can be marked unavailable, and whether the feature can be reproduced from time-appropriate data.
No. Historical repair can make a dataset more useful for research, but it can also give the model information the live system never actually had at decision time.
Because one dropped sequence can create a fictional book state. The resulting metric may look calm, exact, and fully numerical while describing a market that never existed.