Analytics field card

alextouvras.com

Analytics is a stack, not a dashboard.

It’s the system that turns business questions into trusted decisions.

Can we trust this number enough to act on it?

Better dataBetter decisions

The analytics stack

From question to decision

Each layer adds structure, quality and context. Together they turn raw data into trusted insight.

Ask → Grain → Truth → Use

Trust through

Quality, governance and reconciliation

When reports disagree, the system detects, investigates and resolves — before the number reaches the decision. Definitions travel. Reports compose.

Report A4.8%ECL ratio
Report B6.2%ECL ratio

Reconciliation gate

  • Compare and validate
  • Check rules
  • Trace lineage
  • Resolve or block
Trusted result5.1%ECL ratio
Published
  • SecurityWho can see and change the number
  • GovernanceA named owner approves the definition
  • ObservabilityLineage from source to the published measure
  • EvalsTests on gold before a number ships
  • Human ApproveUnexplained gaps stay dark

The 4.8%, 6.2% and 5.1% figures are a picture of a disagreement, not a calculated portfolio result.

Real-world example

Credit risk analytics

Whether to offer an open invoice more time before write-off.

Illustrative. These figures are supposed, not measured.

  1. 01 Ask

    Whether to offer an open invoice more time before write-off. What one row means comes next.

  2. 02 Grain

    Define the reporting unit, reporting date, portfolio scope and treatment of duplicates.

  3. 03 Truth

    Apply approved PD, LGD and EAD definitions, validate the calculations and reconcile totals.

  4. 04 Use

    Expose the validated segment-level drivers in a scorecard that supports investigation and action.

Last month, 100 of 500 issued invoices were overdue. The unpaid amount was 200,000. This month, 112 of 600. The count is 12 percent higher, invoices issued are 20 percent higher, the share is smaller, and the unpaid amount is 280,000. The count, the share, and the money do not agree. Look next at whether a few invoices carry the extra money. That is not yet a conclusion that collections got worse.

Decision
Offer more time before write-off. Whom to call is a different use.
Outcome
The controller can make the offer under the policy. A lower count is not that outcome.
Actor
The credit controller.
If it rises
Check the count, the share, and the unpaid amount. Not an order to refuse time.
Send it
The list can go out as that answer. Looking and watching do not need this line.

When each trial is a slow PD or ECL backtest, that search lives on the Bayesian optimisation card. Grid the cheap notebook. Model the expensive pack.

Architecture

The building blocks

  1. Consumption and visualizationReports, dashboards, apps
  2. Semantic layerMetrics, definitions, logic
  3. Data processingTransform, enrich, validate
  4. Data sourcesInternal and external

Tools and technologies

The right tools for the job

  • Data platform. Databricks, DuckDB, MotherDuck. Snowflake sits in the same lane when the warehouse is the constraint.
  • Transformation. dbt, Python, Jupyter. Spark when the volume requires it.
  • Semantic and BI. Semantic layer, Power BI, Looker. Tableau when that is the room’s consume tool.
  • Orchestration. Airflow or Fabric schedules the path. It does not define the number.
  • Monitoring. Great Expectations for tests. OpenLineage for where the number came from.

Common anti-patterns

Avoid these traps

  • Single source of truth with no reconciliation.
  • Hidden definitions that mean different things.
  • Unvalidated data and no quality checks.
  • Dashboard-only thinking with no decision.
  • No lineage and no way to trace the number.

Maturity ladder

Build over time

  1. 1Ad hocManual, inconsistent, low trust.
  2. 2DefinedStandards and basic governance.
  3. 3ManagedAutomated, monitored, reconciled.
  4. 4OptimizedPredictive, real-time, self-service.

If you can’t name the grain, the owner, and how to stop a bad publish, you don’t ship the number.