Data Provenance Matters More Than Fancy Features

In quant and AI research meetings, feature ideas multiply quickly. Sentiment scores, alternative data, intricate transformations—each promises a new angle. I care about features. I care more about provenance: where the data comes from, how it is revised, who can change it, what latency it carries, and how it fails. Fancy features on shaky ground produce confident nonsense.

Garbage in, governance out

A model can be mathematically elegant and still be a governance failure if nobody can trace inputs. Provenance is how you audit. It is how you debug a sudden performance break. It is how you explain yourself when a strategy misfires. Without it, “the model said so” becomes a shrug.

Leakage deserves special paranoia. Future information sneaking into training is a classic way to manufacture backtest beauty. Provenance discipline makes leakage harder to ignore.

Vendor concentration is research risk

If a desk depends on a single feed for critical decisions, that vendor is part of the strategy’s risk. Map critical paths. Keep an ugly manual fallback when possible. Resilience is not elegance.

Digital-asset data needs extra scepticism

Venue fragmentation, wash-trading suspicions, and inconsistent timestamps can poison conclusions. Treating every on-chain or exchange metric as gospel is naïve. Treating all of it as useless is lazy. Provenance and cross-checks sit in the middle.

What I reward in research culture

I reward researchers who document field definitions, join keys, known quality issues, and revision policies. I am less impressed by a new feature that cannot survive a quality questionnaire. Slow documentation is part of speed—because it prevents rebuilding under panic.

AI systems amplify whatever you feed them. Feeding them well is not glamorous. It is the work.

A simple gate

Contracts and schemas as risk documents

Treat data contracts like investment mandates: definitions, units, calendars, known gaps, and change-notification duties. When a vendor alters a field quietly, strategies can break loudly. Require notice. Test for silent schema change.

Version datasets used in research promotions. If you cannot recreate the exact training inputs, you cannot audit the result. Reproducibility is not academic fetish; it is operational self-defence.

Organisation of ownership

Every critical feed needs an owner on your side, not only a vendor relationship manager. Ownership means knowing the backup plan and the degradation mode. In incidents, minutes matter. Ambiguous ownership burns minutes.

Feature stores and ML platforms can help—or can obscure provenance behind abstraction. Demand lineage views that a sceptical risk person can understand without a week of onboarding.

Fancy features are optional. Knowing what your numbers mean is not.

No feature graduates to live capital without a provenance note and a monitoring plan. If that gate feels annoying, the feature was not ready. Capital prefers boredom to surprise.

Onboarding researchers into provenance culture

New joiners should spend time with data owners before feature brainstorming. Ritualise it. Culture sticks when the first week teaches what “good” looks like. If the first week is only model cleverness, provenance becomes forever secondary.

Praise boring catch reports: someone noticed a silent unit change and prevented a bad promotion. Those stories teach more than another leaderboard of experimental alphas.

Procurement is research quality

Data purchasing decisions are research decisions. Cheap feeds with unclear lineage can cost more than premium feeds with audit trails once a bad promotion lands. Involve risk and engineering in vendor choice, not only research appetite.

When building internal pipelines for Indonesian market education or AI quant work, document lineage from day one. Retrofits are miserable and incomplete. Provenance delayed is provenance denied.

Features impress. Provenance protects. Prioritise protection.

Procurement is research quality

Data purchasing decisions are research decisions. Cheap feeds with unclear lineage can cost more than premium feeds with audit trails once a bad promotion lands. Involve risk and engineering in vendor choice, not only research appetite.

When building internal pipelines for Indonesian market education or AI quant work, document lineage from day one. Retrofits are miserable and incomplete. Provenance delayed is provenance denied.

Features impress. Provenance protects. Prioritise protection.

Working notes I keep returning to

On Data Provenance Matters More Than Fancy Features, I keep a short list of non-negotiables: write the claim, name the failure, size for survival, and review on a cadence that does not depend on mood. The list is not clever. It is how personal opinion stays accountable to capital.

Markets will keep inventing new costumes for old risks. Process language travels better than costume language. That is why I write these notes in public English for anyone building discipline alongside ambition—including students and builders in Indonesia’s growing fintech conversation.

If you change only one behaviour after reading, make it precommitment: decide the rule before the tape demands an exception. Exceptions can still exist. They should expire in writing.

Chase fewer magic variables. Chase cleaner truth about the variables you already claim to trust.