Topic 4 of 20 · Part I : Shaping Model Behaviour

Synthetic data

Synthetic data fills gaps in training and evaluation : but only when it is designed from a coverage plan rather than produced as a large undifferentiated batch.

The control pattern

Data asset

Generated training and test examples

Knobs

Generator model, Constraints, Coverage plan, Rarity, Perturbation level, Human review rate

Outcome

Coverage of important cases that real data does not supply reliably

Measurement

See how to measure this below

What it is

Synthetic data uses generated examples to fill gaps in training and evaluation. It is most useful when real examples are scarce, expensive to label, privacy-sensitive or badly imbalanced.

The generator model and its instructions are knobs, but so are the distribution of cases, the realism of document structure, the frequency of rare events, and the amount of deliberate corruption. Synthetic data should be designed from a coverage plan. Produced as an undifferentiated batch, it amplifies the generator's habits and creates the illusion of diversity without adding genuinely new situations.

Every synthetic record should carry provenance, generation settings, intended difficulty and validation status.

The knobs in detail

Each row is one adjustable property of the data asset, and what moving it tends to do.

KnobWhat you adjustLikely effect
Generator modelWhich model produces the examplesCaps realism and imprints its own habits
ConstraintsRules the generated case must satisfyKeeps examples internally consistent
Coverage planWhich scenarios are represented, and how oftenPrevents illusory diversity
RarityFrequency of unusual eventsTeaches the tail without swamping the centre
Perturbation levelDeliberate corruption and noiseBuilds robustness to real-world mess
Human review rateShare of records checked by a personFocuses scarce review on high-risk labels

Applied: Stocks Assistant

For Stocks Assistant, synthetic cases could model unusual earnings scenarios: a company withdraws guidance, changes reporting segments, announces a restatement, reports a large one-time impairment, changes currency presentation, or gives contradictory signals between prepared remarks and Q&A. Synthetic documents can also reproduce OCR problems, missing table borders, merged columns, low-resolution scans and ticker ambiguity. These are valuable for teaching extraction, uncertainty detection and escalation.

How to measure it

Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.

  • Improvement measured on a separate set of real, point-in-time examples
  • Automated checks on arithmetic, schema validity and absence of future information
  • Reviewer agreement on high-risk labels
  • Coverage achieved against the original plan

Common mistakes

Gains that appear only on synthetic tests. That is a sign the data is teaching artifacts rather than transferable skill.

Generating market outcomes and presenting them as real observations. Synthetic evidence has no place in return prediction.