Home/Tools/Driver Analysis
Customer Experience

Driver Analysis — which two or three things should we improve first to move the metric?

Identifies which attributes most influence a key business outcome — NPS, CSAT, intent to repurchase, churn. Crowdmines fits up to a dozen models in parallel, compares them on cross-validated performance, and produces a defensible importance ranking — with optional SHAP explanations for tree models. For NPS, a two-stage model explains promoter and detractor behaviour separately. The deliverable is a full report — markdown, PDF, PowerPoint and JSON metadata — generated automatically.

SHAP-based importanceTwo-stage NPS modelPowerPoint reportQuadrant prioritisationCross-validated comparison
Driver importance · Telco NPSLive result
What's actually moving NPS?
Top attributes ranked by SHAP-based contribution
Top driver
Network
Top 2
~60%
Models fit
12
Output
PPTX

The business question it answers

“Of everything we measure, which two or three things should we improve first to move our key metric?”

Say your NPS has been flat for three quarters. Feed in your tracker — the NPS column plus 18 attribute ratings (network quality, billing clarity, app usability, store experience, contact centre, etc.). The chat returns: network quality and billing clarity together explain ~60% of the variation in NPS; app usability and contact centre wait times each contribute ~10%; the other 14 attributes are noise. The quadrant chart positions network quality as high importance, low performance— the priority-one investment. Three slides of conclusion, ready for next week’s stakeholder meeting.

How the methodology works

The platform fits up to twelve model families in parallel — tree ensembles (random forest, gradient boosting, XGBoost, LightGBM, CatBoost), linear and logistic regression, partial least squares, MARS, Bayesian ridge, Bayesian networks and a small neural net — then compares them on cross-validated performance and reports which one fits best. For tree models, importance can be reported using SHAP values — a standard technique for quantifying how much each feature contributes to a model’s prediction — when the user opts in; otherwise model-native importance is used.

  • Two-stage NPS model. Separate models for Detractor → Passive and Passive → Promoter, because the things that create promoters are rarely the inverse of the things that create detractors.
  • Quadrant chart. Importance vs. current performance. Anything in the top-left (high importance, low score) is the priority-one investment.
  • Importance tiers. Features are categorised Primary / Secondary / Low Priority using quantile thresholds (75th and 40th percentile of normalised importance).
  • Optional calibration weighting. Inverse- frequency weights correct for over- or under-representation without requiring external population targets; effective sample size is reported alongside.

What you see in the chat

A model-comparison summary (which model performed best and why), an importance ranking — SHAP-based for tree models when enabled, model-native otherwise — the quadrant map for prioritisation and the two-stage NPS breakdown when applicable. The complete report is downloadable in markdown, PDF, PowerPoint and JSON metadata — the auto-generated PowerPoint is the deliverable that closes the gap between running the analysis and having something to present on Monday.

Required data
Dependent varNPS · CSAT · churn · intent
Independent varsAttribute ratings, demographics
DV schemeNPS · Top-2-box · custom · none
Min sampleA few hundred per outcome
OptionalCalibration weighting · control vars
Use when you hear
  • What should we improve first to lift NPS?
  • We measure 30 things in our tracker — which ones actually matter?
  • Our experience team is trying to prioritise investments and needs evidence.
  • Why is NPS dropping? / What’s behind the satisfaction lift in Q2?
  • We need a defensible story for the board on what’s moving the metric.
Stakeholder-ready

The auto-generated PPTX

The deliverable that closes the gap between running the analysis and having something to present on Monday — model comparison, importance rankings, quadrant maps and an executive summary, all in a slide deck you can take straight into a stakeholder meeting.

Step by step

From tracker upload to PowerPoint export.

Driver Analysis has a richer interactive flow than the simpler tools — research-question framing, variable mapping, DV recoding, and model selection are all guided by the chat.

1
Ask the question
“What’s driving NPS in our business banking segment?” — the platform recognises the intent and selects Driver Analysis.
2
State the research question
Articulate the specific business question (“What drives NPS among our premium customers?”). It threads through the entire analysis and the final report.
3
Map the variables
Identify the dependent variable(s) and the independent attributes. Mappings are suggested from column names and types.
4
Pick the DV scheme
NPS / Top-2-box / custom / none — auto-suggested based on the variable type. NPS triggers the two-stage model.
5
Review the model plan
The platform recommends which model families to fit based on DV type and sample size. Add or remove models in the chat.
6
Models run in parallel
Cross-validated model comparison; SHAP importance for tree models when enabled, model-native otherwise.
7
Download the report
Model comparison, importance ranking, quadrant map, two-stage NPS breakdown, executive summary — markdown, PDF, PowerPoint and JSON metadata.
Compared to

How Crowdmines compares to SPSS, specialist tools and consulting agencies.

Driver analysis (also called key-driver, importance-performance, or relative importance analysis) is a staple of CX, brand and product research. Traditional approaches range from manual regression in SPSS / SAS / R, to specialised tools (Displayr, Q Research), to consulting engagements where an agency runs the analysis and delivers a deck two weeks later.

CapabilityTraditional (SPSS / R / SAS)Specialised tools (Displayr, Q Research)Agency / consultingCrowdmines
Setup effortWrite regression syntax, manually select models, configure SHAP librariesPoint-and-click, but still requires variable selection and model configurationBrief the agency, send data, waitAsk a question in the chat
Number of modelsTypically one model per run; analyst picks whichUsually one or two model typesAgency picks a model (often just regression)Up to a dozen model families fit in parallel, auto-compared
SHAP explanationsRequires separate Python library, custom codeSome tools support it; many don'tRarely included — agencies prefer simpler methodsAvailable for all five tree ensembles when enabled
Control variablesAnalyst hand-codes the partiallingRare — most tools don't partial out controlsSometimes includedBuilt-in — controls partialled out, reported drivers renormalised
NPS two-stageMust manually split data, run twice, combineRare — most tools treat NPS as a single outcomeSome agencies do this; many don'tAutomatic when NPS scheme is selected
Quadrant mappingAnalyst builds manually in Excel or ggplotSome tools include itDelivered in the agency's deckAuto-generated, interactive
Calibration impactRun weighted + unweighted, diff the rankings by handRareRareBuilt-in weighted-vs-unweighted comparison with a rank-change table and AI narrative
Report generationAnalyst builds slides manually (hours to days)In-platform view; export is limitedAgency delivers a deck (days to weeks)Auto-generated markdown, PDF, PowerPoint and JSON metadata in minutes
Cross-validationAnalyst must code itSome tools support itRarely transparent to the clientUp to 5-fold CV across every fitted model, reported automatically
TurnaroundHours to daysHoursDays to weeksMinutes
Beta Program Open

Three slides of conclusion — ready for next week’s stakeholder meeting.

Built for CX, insights and brand teams who need a defensible story for the next stakeholder meeting. Multi-model comparison, SHAP importance, two-stage NPS, quadrant map — and a presentation-ready PowerPoint.