Day 1 lecture: Data-driven decisions¶

Organisations make better decisions when they combine data, analysis and judgement. This lecture maps the course, explains the four kinds of analytics, sets out how an analysis is organised, and shows live why summary numbers alone can mislead.

Open this notebook in JupyterLab from the course folder and run each cell from top to bottom with Shift + Enter.

What you will be able to do¶

1. Find your way around the course: its outcomes, graded work and schedule.
2. Explain why data-driven decisions tend to beat decisions made on instinct alone, and where judgement still matters.
3. Tell descriptive, diagnostic, predictive and prescriptive analytics apart, with a business example of each.
4. Follow the path from question to data, analysis, validation and decision, and turn a vague request into an answerable question.
5. Show with code why summary statistics are not enough, using Anscombe's quartet and the Datasaurus Dozen.
6. Name the most common ways numbers mislead.

Reading: Gordon (2023), Business analytics, Chapter 1 (how data and analytics can improve human decisions) and Chapter 4 (analysis fundamentals).

How to use the questions in this notebook

Predict, then reveal boxes ask you to guess before running the next cell. Write your guess down first: being wrong is how you learn what to check. Check your understanding boxes are short questions to answer in your own words, alone or with the person next to you.

Part 1: The course map¶

DATA 300 Data Driven Decision Making moves from describing what happened, to predicting what will happen, to deciding what to do. You will work in Python with Jupyter, Tableau and Excel.

Course learning outcomes (CLOs)¶

Every graded item is tied to one or more of these outcomes.

How the course is organised¶

Graded work¶

All work is submitted as a PDF on GAP (Moodle).

How the days work¶

Mornings are theory, like this lecture. Most afternoons continue the theory and end with a lab, where you complete a graded item in class. The schedule lists what happens each day and the deadlines; the setup page explains the software.

Before today, in the pre-class week, you set up your computer, worked through Python and Jupyter basics and submitted Activity 1, the readiness check. Today:

Check your understanding

Look at the graded-work table. Which single item is worth the most? Which outcome (CLO) do you expect today's work to support most?

Part 2: Why data-driven decisions?¶

Every organisation decides things all the time: what to stock, whom to hire, which customers to target, how much to charge. Many of those decisions are still made on gut feeling, habit or the opinion of the most senior person in the room.

The idea in one line¶

The course textbook's subtitle sums it up: good decisions combine data, analysis and judgement. Data on its own decides nothing. Analysis turns data into evidence. Judgement, from people who know the business, decides what the evidence means and what to do.

Why instinct alone is risky¶

Human judgement is valuable, but it has predictable weak spots. Psychologists have named many of them:

Data and analysis do not remove these biases, but they give everyone something to check against. A decision based on evidence can also be explained, repeated and improved, because you can see what it rested on.

Where judgement still matters¶

Analytics supports decisions; it does not make them. People still have to:

• choose the question worth answering;
• judge whether the data is trustworthy and relevant;
• weigh things the data does not capture: ethics, strategy, staff morale, reputation;
• take responsibility for the decision.

Stories where data changed the outcome¶

The full stories are in Why data analytics matters.

Predict, then reveal

Four of these stories are successes; one is a failure. Before reading the last row, ask yourself: what could make a model that once worked stop working? Then read the last row.

Check your understanding

Think of one decision at a workplace or organisation you know that was made on gut feeling. What data could have informed it, and which of the four weak spots might have affected it?

Part 3: Four kinds of analytics¶

Analytics answers four kinds of question. Each builds on the one before.

Many textbooks, including ours, group the first two together and speak of descriptive, predictive and prescriptive analytics. Diagnostic work is the bridge: it explains the description before anyone predicts or prescribes.

Across the business¶

Where each kind appears in this course¶

A small live example¶

The cell below loads the table of restaurant bills you met in the pre-class Python and Jupyter basics notebook. Each row is one table's bill and the tip it left (in US dollars, from a real restaurant study). We use it only to see the first three kinds of question in code.

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd

tips = pd.read_csv("data/tips.csv")
tips["tip_percent"] = 100 * tips["tip"] / tips["total_bill"]

# Descriptive: what happened? Average tip on each day.
print("Descriptive: average tip (US dollars) by day")
print(tips.groupby("day")["tip"].mean().round(2))

Descriptive: average tip (US dollars) by day
day
Fri     2.73
Sat     2.99
Sun     3.26
Thur    2.77
Name: tip, dtype: float64

Predict, then reveal

One day has the highest average tip. Is that because people are more generous that day, or because they spend more? Write your guess, then run the next cell, which looks at the average bill and the tip as a percentage of the bill.

# Diagnostic: why? Compare the bill size and the tip as a share of the bill.
print("Diagnostic: average bill and tip percentage by day")
tips.groupby("day").agg(
    average_bill=("total_bill", "mean"),
    average_tip_percent=("tip_percent", "mean"),
).round(2)

Diagnostic: average bill and tip percentage by day

Compare the two columns in the output above. Where the average tip is high, look at whether the bill or the percentage is what differs. A difference in amounts can come from a difference in spending rather than in generosity: that is a diagnostic finding.

Now a prediction: the straight line that best fits tip against bill lets us estimate the tip on a new bill.

# Predictive: what is likely? Fit a straight line of tip against bill.
slope, intercept = np.polyfit(tips["total_bill"], tips["tip"], 1)
new_bill = 40
print(f"Best-fit line: tip = {intercept:.2f} + {slope:.3f} x bill")
print(f"Predicted tip on a {new_bill}-dollar bill: {intercept + slope * new_bill:.2f} dollars")

Best-fit line: tip = 0.92 + 0.105 x bill
Predicted tip on a 40-dollar bill: 5.12 dollars

Prescriptive analytics would go one step further: compare options (for example, which evening to add a second server) against an objective (service quality, wages) under constraints (staff available), and recommend one. You will build that kind of model in Excel on Day 5.

Check your understanding

Which kind of analytics is each statement?

1. "Deliveries in the North got slower because the new carrier skips Saturday collections."
2. "We expect about 400 returns in December."
3. "Move the slow carrier's routes to the faster one for one month and measure the effect."
4. "We shipped 12,000 parcels in October."
5. "A customer who has not ordered for 90 days is likely to leave."

Part 4: Analysis fundamentals — from question to decision¶

A useful analysis supports a decision. It follows the same path every time:

Question → Data → Analysis → Validation → Decision

The course site describes the same path as seven steps, splitting out preparation and visualisation.

Step 1 in detail: a question you can answer¶

A manager says: "Sales feel uneven across our customer groups." That is a feeling, not a question. A question you can answer with data names four things:

"Which customer segment had the highest total sales over the last three years, so we can choose one segment to investigate next month?"

Step 2 in detail: what does one row mean?¶

A manager asks: "How many orders did we have?" The sales file has one row per order line, so an order with three products takes three rows. Counting rows counts lines, not orders. You must first know what one row represents (the table's grain), then count distinct order numbers. You practised this in the pre-class Python and Jupyter basics notebook.

Step 4 in detail: validation¶

A plausible number is not a checked number

Before you report a result: do the parts add up to the total? Does a count match a simple manual check? Can you explain the largest and smallest values? Have you looked at the data in a chart, not only at summary numbers? Parts 5 and 6 show why that last check matters.

Check your understanding

A colleague asks: "Is the new website working?" Which of the four parts (decision, population, period, measure) are missing? Suggest one value for each.

Part 5: Statistics are not enough — Anscombe's quartet¶

In 1973 the statistician Francis Anscombe made four small datasets to make one point. We build them in code, compute their summary statistics, and only then look at them.

The code stores all four datasets in one table. Each row is one point: which set it belongs to (set) and its x and y values.

# Anscombe's quartet (1973): four small datasets, eleven points each.
x = [10, 8, 13, 9, 11, 14, 6, 4, 12, 7, 5]
anscombe = pd.concat([
    pd.DataFrame({"set": "I",   "x": x, "y": [8.04, 6.95, 7.58, 8.81, 8.33, 9.96, 7.24, 4.26, 10.84, 4.82, 5.68]}),
    pd.DataFrame({"set": "II",  "x": x, "y": [9.14, 8.14, 8.74, 8.77, 9.26, 8.10, 6.13, 3.10, 9.13, 7.26, 4.74]}),
    pd.DataFrame({"set": "III", "x": x, "y": [7.46, 6.77, 12.74, 7.11, 7.81, 8.84, 6.08, 5.39, 8.15, 6.42, 5.73]}),
    pd.DataFrame({"set": "IV",  "x": [8, 8, 8, 8, 8, 8, 8, 19, 8, 8, 8],
                  "y": [6.58, 5.76, 7.71, 8.84, 8.47, 7.04, 5.25, 12.50, 5.56, 7.91, 6.89]}),
], ignore_index=True)

print("Rows, columns:", anscombe.shape)
anscombe.head()

Rows, columns: (44, 3)

Now the summary statistics for each set. groupby("set") splits the table into the four sets; agg computes several summaries for each one. The correlation measures how closely the points follow a straight line, from −1 to +1.

summary = anscombe.groupby("set").agg(
    mean_x=("x", "mean"),
    mean_y=("y", "mean"),
    sd_x=("x", "std"),
    sd_y=("y", "std"),
)
summary["correlation"] = anscombe.groupby("set")[["x", "y"]].apply(lambda d: d["x"].corr(d["y"]))
summary.round(2)

Finally, the straight line that best fits each set. np.polyfit(x, y, 1) returns its slope and intercept.

for name, points in anscombe.groupby("set"):
    slope, intercept = np.polyfit(points["x"], points["y"], 1)
    print(f"Set {name:>3}: best-fit line  y = {intercept:.2f} + {slope:.3f} x")

Set   I: best-fit line  y = 3.00 + 0.500 x
Set  II: best-fit line  y = 3.00 + 0.500 x
Set III: best-fit line  y = 3.00 + 0.500 x
Set  IV: best-fit line  y = 3.00 + 0.500 x

Predict, then reveal

Read the outputs above. The four sets share their means, spreads, correlation and best-fit line. Sketch what you think a scatter plot of each set looks like. Will the four pictures be similar? Then run the next cell.

fig, axes = plt.subplots(2, 2, figsize=(8, 6.5), sharex=True, sharey=True)
line_x = np.array([2, 20])
for ax, (name, points) in zip(axes.flat, anscombe.groupby("set")):
    slope, intercept = np.polyfit(points["x"], points["y"], 1)
    ax.plot(line_x, intercept + slope * line_x, color="#eb6834", linewidth=2)
    ax.scatter(points["x"], points["y"], s=40, color="#2a78d6", zorder=3)
    ax.set_title(f"Set {name}")
for ax in axes[1]:
    ax.set_xlabel("x (no units)")
for ax in axes[:, 0]:
    ax.set_ylabel("y (no units)")
fig.suptitle("Same statistics, same trend line, four different stories", x=0.01, ha="left", fontweight="bold")
fig.tight_layout()
plt.show()

Remember

Summary statistics compress data, and compression throws information away. A chart shows you what was thrown away. Plot first, then summarise.

Check your understanding

In set IV, suppose the one point far to the right is a data-entry error and you remove it. What happens to the correlation? Why can you not even compute it?

Part 6: The Datasaurus Dozen¶

In 2017 Justin Matejka and George Fitzmaurice pushed Anscombe's idea further. Starting from a dinosaur drawn with data points (originally by Alberto Cairo), they nudged the points step by step into twelve other shapes while keeping the summary statistics almost identical.

Their data is in the course folder, so it loads without internet.

# The Datasaurus Dozen (Matejka and Fitzmaurice, 2017), from the open-source
# datasauRus package (MIT licence), bundled in the data folder.
dozen = pd.read_csv("data/datasaurus_dozen.csv")

print("Rows, columns:", dozen.shape)
print("Datasets:", dozen["dataset"].nunique())
dozen.head()

Rows, columns: (1846, 3)
Datasets: 13

Predict, then reveal

The table has thirteen datasets. Their names (dino, star, circle, …) are a hint. First look at the statistics in the next cell. Would you have guessed the datasets had different shapes from these numbers alone?

dozen_summary = dozen.groupby("dataset").agg(
    mean_x=("x", "mean"),
    mean_y=("y", "mean"),
    sd_x=("x", "std"),
    sd_y=("y", "std"),
)
dozen_summary["correlation"] = dozen.groupby("dataset")[["x", "y"]].apply(lambda d: d["x"].corr(d["y"]))
dozen_summary.round(2)

Now draw one small scatter plot per dataset. Charts like this, one per group on the same axes, are called small multiples.

names = dozen["dataset"].unique()
fig, axes = plt.subplots(3, 5, figsize=(12, 7), sharex=True, sharey=True)
for ax in axes.flat:
    ax.set_visible(False)
for ax, name in zip(axes.flat, names):
    points = dozen[dozen["dataset"] == name]
    ax.set_visible(True)
    ax.scatter(points["x"], points["y"], s=8, color="#2a78d6")
    ax.set_title(name.replace("_", " "))
fig.suptitle("The Datasaurus Dozen: same statistics, thirteen shapes (x and y have no units)",
             x=0.01, ha="left", fontweight="bold")
fig.tight_layout()
plt.show()

A report saying "x and y are essentially unrelated, centred around the same values" would be true for every panel, including the dinosaur.

At work

You will rarely meet a dinosaur in a sales report. You will meet the same problem in everyday forms: an average delivery time that hides a group of customers waiting weeks; one huge order that makes a quiet month look like growth; two stores with the same average sales but very different patterns. See Statistics are not enough.

Check your understanding

Why are the thirteen panels drawn with the same axes? What could go wrong if each panel chose its own axis range?

Part 7: A preview of how numbers mislead¶

Numbers feel objective, but the same data can tell very different stories depending on how it is summarised and drawn. Here is one trap, live. Both charts below show the same two numbers.

Predict, then reveal

Two products have satisfaction scores of 96% and 98%. Before running the cell, decide: is that a big difference or a small one? Then see how each chart makes you feel about it.

products = ["Product A", "Product B"]
satisfaction = [96, 98]  # made-up percentages

fig, (left, right) = plt.subplots(1, 2, figsize=(9, 3.5))
left.bar(products, satisfaction, color="#2a78d6")
left.set_ylim(95, 98.5)
left.set_title("Axis starts at 95")
right.bar(products, satisfaction, color="#2a78d6")
right.set_ylim(0, 100)
right.set_title("Axis starts at 0")
for ax in (left, right):
    ax.set_ylabel("Satisfied customers (%)")
fig.suptitle("Same data, two impressions (made-up numbers)", x=0.01, ha="left", fontweight="bold")
fig.tight_layout()
plt.show()

We judge bars by their length, so bar charts should start at zero. On the left, Product B looks several times better; on the right you see two products that are almost the same.

The full list of traps is in How statistics can mislead. We return to each one later in the course.

A checklist for any number you are shown

1. Compared to what?
2. Who or what is included?
3. What is the denominator?
4. Could something else explain it?
5. How is it drawn?

Summary¶

• Good decisions combine data, analysis and judgement. Data gives everyone something to check instinct against; people still choose the question and own the decision.
• Analytics answers four kinds of question: what happened (descriptive), why (diagnostic), what next (predictive) and what to do (prescriptive).
• Every analysis runs question → data → analysis → validation → decision. A good question names a decision, population, period and measure.
• Different data can share identical summary statistics. Always plot your data.
• Numbers mislead through causation claims, mixed groups, axes, time windows, averages, missing bases and survivors.

Closing quiz¶

Answer in your own words, alone, in two or three minutes per question. We discuss the answers before lunch.

1. Name the four kinds of analytics and give a one-line business example of the predictive kind.
2. Give one reason why a decision based only on a manager's instinct can go wrong, and one thing data cannot decide for you.
3. Rewrite "How are our stores doing?" as an analytical question that names a decision, population, period and measure.
4. Two datasets have the same mean, standard deviation and correlation. Can you conclude they look alike? What should you do?
5. In Anscombe's set III, what single feature of the data changes the best-fit line?
6. Why should a bar chart's value axis start at zero?
7. In the path question → data → analysis → validation → decision, what is the difference between analysis and validation?

This afternoon¶

1. Theory: Exploring data in Python: structure and data quality, descriptive statistics and skew, pivot tables, correlation, confounding and charts for large datasets.
2. Lab: Applied Lab 1: Data exploration (3%), where you apply the same steps to a table of more than 50,000 diamonds.

Bring a working JupyterLab: if the setup check did not pass on your laptop, tell your facilitator at lunch.