Descriptive Pattern Analysis

Descriptive Pattern Analysis

This skill covers descriptive, already-known-dimension analysis — the step between understanding a dataset (data-profiling) and communicating findings (analytics-storytelling) for requests that aren't actual-vs-plan variance, cross-system reconciliation, or model-building. It exists because a real trend claim, a real correlation claim, and a real benchmark claim each carry a specific statistical meaning that is easy to conflate with an informal one (a YTD total, a coincidence, an arbitrary reference point) — and a finance/accounting audience has no independent way to catch the difference.

When to Use

• Describing a trend over time (in Hyndman's sense: a long-term direction net of seasonal/cyclic effects, not a raw running total or plain line chart).
• Comparing an already-known cohort or category — region, vendor tier, cost center, business unit — against another, or against itself over time.
• Investigating whether two or more measured variables move together (e.g., throughput vs. maintenance capex), stated as a relationship, not a cause.
• Benchmarking a metric against an internal or external reference point.
• Following up on an anomaly data-profiling flagged, to determine whether it's a real business signal or a data quality issue.

Relationship to Other Skills

• Use data-profiling first to understand the raw dataset; this skill picks up once the data is known to be usable.
• Use predictive-model-development instead when the grouping itself is unknown and must be discovered (clustering), when a model is being built, or when a genuinely uncertain future value must be forecast.
• Use financial-variance-analysis instead for actual-vs-plan or period-over-period variance against predefined dimensions.
• Use reconciliation-analytics instead for cross-system or ledger matching.
• Use analytics-storytelling for the general principles of narrative communication; Stage 4 below applies those principles specifically to a descriptive finding, but does not replace that skill for other deliverables.

Workflow

1. Frame the descriptive question

   • State the pattern being investigated, the known dimension or time window it applies to, and what it's being compared against (a prior period, a peer, an internal reference, or nothing — a pure observation).
   • Confirm the dimension or grouping is already known and defined; if it isn't, this is a predictive-model-development request instead.

2. Select and apply the technique

   • Choose from: trend (decomposition-aware — separate direction from seasonal/cyclic effects before claiming a trend exists), cohort/category comparison, correlation between measured variables, or benchmarking against a reference.
   • State briefly which technique was used and why it fits the question — this doesn't need the full rejected-alternatives log predictive-model-development requires, but the choice should not be silent.

3. Check materiality and robustness

   • Confirm the pattern is large enough to matter, not just visible.
   • Confirm the sample size or time window is adequate to support the claim.
   • Check whether removing an outlier or extending the window changes the conclusion.
   • For correlation claims, check whether both series share a common trend and re-test on detrended data (or period-over-period deltas) before concluding the relationship is real — two unrelated trending series will show high correlation purely because both are moving over time.

4. Translate for a finance/accounting audience

   • State the finding in plain language before any methodology.
   • If the finding involves correlation, state it as a relationship, not a cause, unless a controlled comparison supports causation.
   • Express uncertainty honestly rather than with false precision.
   • Avoid unexplained statistical jargon; if a technical term is necessary, define it in one clause.
   • Close with a specific, actionable recommendation tied to the finding.

Output Template

# Descriptive Analysis: [Question or Metric]

## 1. Framing
- Pattern being investigated:
- Known dimension / time window:
- Compared against:

## 2. Technique
- Technique applied:
- Why it fits this question:

## 3. Materiality & Robustness
- Is the pattern large enough to matter:
- Sample size / window adequacy:
- Sensitivity to outliers or window changes:

## 4. Translation
- Plain-language finding:
- Relationship vs. causation (if applicable):
- Uncertainty:
- Recommended action:

Quality Checklist

• A trend claim reflects a decomposed long-term direction, not a raw cumulative total or an unexamined line chart.
• Cohort/category comparisons use dimensions that were already known going in, not an undisclosed discovered grouping.
• Correlation is never presented as causation without a controlled comparison to support it.
• A correlation between two trending series is re-tested on detrended data before being called real, not accepted at face value.
• The benchmark or comparison reference is explicit and defensible, not arbitrary.
• Materiality and robustness are checked before a finding is presented as decision-worthy.
• Translation is jargon-free and states uncertainty honestly.