Storytelling with Data

Context checklist before proceeding:

[ ] I can name my primary audience
[ ] I have written a one-sentence Big Idea
[ ] I know what action I want them to take
[ ] I know if this is a live presentation, document, or dashboard

Consult references/principles/context-setting.md for the full audience analysis worksheet.

Step 2 — Choose an Appropriate Display

Use the correct chart for the relationship in your data. Never pick a chart because it looks impressive.

Rule: If you are using a pie chart, ask whether a bar chart would be clearer. The answer is almost always yes.

Consult references/principles/chart-selection.md for full decision guide with examples.

Step 3 — Eliminate Clutter

Clutter is any visual element that does not add informational value. It increases cognitive load without increasing understanding.

Elements to eliminate by default:

• Chart borders and frames
• Gridlines (remove or lighten to near-invisible)
• Data markers on every point of a line
• Axis labels that duplicate data labels
• Legends when you can label directly
• 3D effects, shadows, gradients
• Dual y-axes (almost always)
• Titles that describe what ("Sales by Month") instead of the insight ("Sales peaked in Q3")

The data-ink test: For every element, ask "Would removing this lose information?" If no → remove it.

# Databricks: standard clutter elimination
import matplotlib.pyplot as plt

def declutter(ax):
    """Remove visual noise from a matplotlib axes."""
    ax.spines['top'].set_visible(False)
    ax.spines['right'].set_visible(False)
    ax.spines['left'].set_color('#CCCCCC')
    ax.spines['bottom'].set_color('#CCCCCC')
    ax.tick_params(colors='#666666')
    ax.yaxis.grid(True, color='#EEEEEE', linewidth=0.8, zorder=0)
    ax.set_axisbelow(True)
    return ax

Consult references/principles/clutter-elimination.md for the full declutter checklist.

Step 4 — Focus Attention

The audience's eye goes where you direct it. Use pre-attentive attributes to highlight the signal and suppress the noise.

Pre-attentive attributes (processed in < 250ms, before conscious thought):

Golden rule: Use one accent color. Everything else is gray (#CCCCCC or #AAAAAA). The accent draws the eye; gray provides context without competing.

# Storytelling with Data color pattern
GRAY_LIGHT  = '#CCCCCC'
GRAY_MED    = '#888888'
GRAY_DARK   = '#444444'
ACCENT      = '#E8664A'   # warm coral — stands out, not alarming
ACCENT_BLUE = '#1A77B5'   # for positive / neutral emphasis

def apply_swd_palette(ax, highlight_index, bars):
    """Color one bar accent, all others gray."""
    for i, bar in enumerate(bars):
        bar.set_color(ACCENT if i == highlight_index else GRAY_LIGHT)

Consult references/principles/pre-attentive-attributes.md for full visual emphasis guide.

Step 5 — Think Like a Designer

Design is not decoration — it is the arrangement of elements to serve communication.

Affordance: Make it obvious how to read the chart. Labels should sit next to what they label. Axes should have clear titles with units.

Accessibility:

• Colorblind-safe: never use red + green as the only differentiator
• Minimum font size: 10pt for body, 12pt for titles
• All text horizontal (never rotated axis labels — use horizontal bar instead)

Aesthetics — the three rules:

1. Use white space intentionally; do not fill every pixel
2. Align elements on a grid; ragged layouts look unfinished
3. Choose one type family; vary weight and size, not the font

Gestalt principles that apply most to data viz:

Consult references/principles/design-principles.md for the full design checklist.

Step 6 — Tell a Story

A presentation without a narrative is a data dump. Structure your communication like a story.

The Story Arc:

Setup (tension)           → Conflict (complication)      → Resolution (call to action)
"Here is what we expected"  "Here is what actually happened"  "Here is what we should do"

The 3-Minute Story: Distill your entire message to what you would say if you had only 3 minutes. This forces clarity. Then expand.

Annotation as narrative: Use text directly on the chart to guide the audience. Never make them look back and forth between a legend and the chart.

# Annotating the key moment in a line chart
ax.annotate(
    'Policy change\ncaused spike',
    xy=(highlight_x, highlight_y),
    xytext=(highlight_x + 2, highlight_y + 5),
    arrowprops=dict(arrowstyle='->', color=GRAY_DARK, lw=1.2),
    fontsize=9, color=GRAY_DARK,
    ha='left'
)

Slide / section title rule: The title IS the takeaway.

• Bad title: "Revenue by Region Q1–Q4 2024"
• Good title: "Northeast revenue fell 18% in Q3 — driven by supply disruption"

Consult references/principles/narrative-structure.md for story templates, slide structure, and the full annotation guide.

Workflow

When a user brings data to communicate, follow this sequence:

1. Elicit context — Ask who the audience is and what action is needed. Do not proceed without a Big Idea.
2. Audit the current visual (if one exists) — Apply the clutter checklist. List every element that should be removed.
3. Select or confirm chart type — Justify with the chart selection guide.
4. Apply the declutter template — Use declutter(ax) + strip spines.
5. Apply pre-attentive attributes — One accent color; everything else gray.
6. Write the chart title as the insight — Not the description.
7. Add annotations — Mark the moment that supports the Big Idea.
8. Render in Databricks — display(fig) + plt.close(fig).

Databricks Quick Start

import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import numpy as np

# SWD color constants
GRAY_LIGHT  = '#CCCCCC'
GRAY_MED    = '#888888'
ACCENT      = '#E8664A'

# --- Data ---
categories = ['Product A', 'Product B', 'Product C', 'Product D', 'Product E']
values     = [42, 78, 55, 91, 63]
highlight  = 3   # index of the bar we want the audience to focus on

# --- Build chart ---
fig, ax = plt.subplots(figsize=(6, 3.5))

colors = [ACCENT if i == highlight else GRAY_LIGHT for i in range(len(categories))]
bars = ax.barh(categories, values, color=colors, height=0.55)

# Eliminate clutter
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.spines['left'].set_visible(False)
ax.spines['bottom'].set_visible(False)
ax.xaxis.set_visible(False)
ax.tick_params(left=False)

# Direct labels — no axis needed
for bar, val in zip(bars, values):
    ax.text(bar.get_width() + 1, bar.get_y() + bar.get_height() / 2,
            f'{val}%', va='center', fontsize=9,
            color=ACCENT if val == values[highlight] else GRAY_MED,
            fontweight='bold' if val == values[highlight] else 'normal')

# Title IS the insight
ax.set_title('Product D leads with 91% satisfaction — 13 pts above average',
             fontsize=11, fontweight='bold', color='#222222', loc='left', pad=12)

display(fig)
plt.close(fig)

Domain Context: Credit and Risk Analytics

Audience Tiers

The same model or analysis requires a different visualization for each audience. Always identify the primary recipient before choosing a chart type.

A KS curve in an executive slide will confuse; a single summary number in a risk review will be rejected as insufficient. These are not interchangeable.

Trustworthiness in Regulated Environments

Visualizations in formal reports — model validation documents, regulatory submissions, audit trails — carry an obligation of accuracy that goes beyond aesthetic choices.

Non-negotiable rules:

• Bar charts must start at 0 — a cropped y-axis artificially inflates differences
• Dual y-axis must not be used in formal reports — the relationship between two arbitrary scales is misleading
• Sample size and observation period must be stated on every chart
• Trend lines must declare the method (rolling average, linear regression, LOWESS)
• Color must not be the sole information carrier — annotate values directly
• All charts in versioned reports carry a data-as-of date

Task Map by Domain

Credit and Risk Analytics

Model Evaluation (all domains)

Fraud Detection

Customer Analytics

Causal Inference

Common Anti-Patterns and Fixes

Resources

references/principles/

Communication and design fundamentals — apply to all chart types.

• context-setting.md — Audience analysis worksheet, Big Idea template, communication mode guide
• chart-selection.md — Chart decision guide with code templates for each type
• clutter-elimination.md — Declutter checklist, data-ink ratio, before/after examples
• pre-attentive-attributes.md — Pre-attentive attribute reference with code patterns
• design-principles.md — Gestalt principles, accessibility checklist, typography
• narrative-structure.md — Story arc templates, annotation guide, slide title rules

references/domain/

Domain-specific chart patterns with ready-to-run code.

• credit-risk-charts.md — KS curve, vintage, migration matrix, risk heatmap, bullet chart, PSI
• model-evaluation-viz.md — ROC, calibration, SHAP waterfall, feature importance, lift/gain
• fraud-detection-charts.md — Anomaly time series, dual panel, scatter outlier, calendar heatmap, rolling z-score
• customer-analytics-charts.md — Cohort retention, funnel, segment profile, churn distribution, CLV box, A/B test CI
• causal-inference-charts.md — Coefficient plot, parallel trends, DiD event study, RDD binned scatter, propensity overlap

scripts/

• swd_style.py — General: declutter(), apply_swd_palette(), annotate_insight(), insight_title(), label_bars(), highlight_region() Domain: risk_colormap(), psi_status(), fmt_pct(), fmt_bps(), waterfall_colors()