Lesson 8.4: Presenting results, portfolios and interviews
Turn analysis into decisions: answer-first writing, impact in rupees with a range, honest charts, a portfolio that shows your work, and interview practice.
By the end of this lesson you will be able to shape an analysis into a recommendation a busy decision-maker can act on, put a rupee value and an honest range on a model's benefit, choose and label charts so that they carry one message, assemble a portfolio of projects that shows how you work, write resume lines that state results, and prepare for the main kinds of data science interview. You will turn a 0.6-point fall in bad rate into an extra ₹2.77 crore a year with a range of ₹1.39 to ₹4.16 crore, work out that a test needs 7,600 customers in each group, and show that a fraud alert from a very accurate test is right only 8.3 per cent of the time. You should have read Lessons 1.4, 2.3 and 6.3 for charts, intervals and cutoffs, and Lesson 8.3, whose scorecard is the running example. The mathematics is the arithmetic of rates and ranges you already use.
The best model in the world earns nothing until a person decides differently because of it. This is the part of the work that textbooks leave out and workplaces reward. An analysis is finished when the right people understand it, believe it for the right reasons and know what to do next. The same skills carry into your own career: a portfolio is a presentation of your work to strangers, and an interview is a conversation in which you present your thinking live. This lesson treats them together.
Start from the decision, not from the analysis
Most weak presentations narrate the work in the order it was done: the data, the cleaning, the models tried, the metrics, and, at the end, almost as an afterthought, what the person should do. The audience is lost long before the end. A strong presentation inverts the order. It starts with the answer and then supplies only as much evidence as the audience needs to trust it.
Different people need different things from the same project.
Audience · What they need · What to leave outExecutive or business owner · The decision, the value in rupees, the main risk, the ask · Method, code, most metrics
Product or operations manager · What changes in the process, who is affected, how success will be measured · Algorithm details
Risk, compliance or audit · Definitions, data, validation, limits, monitoring, who signed off · Marketing language
Fellow data scientists · Method, validation, code, what you would try next · Business context they know
Write for the most senior reader first and put the rest in an appendix. A useful test: if the reader only read the first paragraph, would they know what to do?
An answer-first structure
A reliable shape for a written summary or a spoken briefing is five parts.
1. The recommendation. One or two sentences: what should be done.
2. The value. The effect in the audience's own units, such as rupees, customers or hours, with a range.
3. The evidence. Two or three facts, charts or numbers that make the recommendation credible.
4. The risks and limits. What could make it wrong, and what you did not test.
5. The ask and the next step. The decision needed, from whom, by when.
Compare two openings for the scorecard of Lesson 8.3. The first: "We explored income, employment type and missed payments, binned them, computed weight of evidence, fitted a logistic regression and obtained a Gini of 0.41." The second: "I recommend we use the new scorecard with a cutoff of 550. On last year's loans it would have earned about ₹31 lakh more per 5,000 applications than approving everyone, by declining the 26 per cent of applicants whose expected loss exceeds their profit. It ranks well on a later sample, with a Gini of 0.38. The main risk is that it rests on 301 defaults, so we propose a three-month trial on 20 per cent of applications and a monitoring plan. We need a decision on the trial by Friday." The second takes thirty seconds to read, and a reader can agree, argue or ask a precise question. The first invites a nod and nothing else.
Put a value on the result, with a range
Executives rarely care about a Gini. They care about money, risk and customers. Your job is to translate. The translation is arithmetic, and it must come with honest uncertainty.
Suppose a lender receives 100,000 applications a year and approves 70 per cent, 70,000 loans. The current process has a bad rate of 4.0 per cent among approved loans, and the new scorecard, at the same approval volume, is expected to bring it to 3.4 per cent, an improvement of 0.6 points. A good loan earns ₹6,000 and a bad one loses ₹60,000. Every bad loan avoided is replaced by a good one, so it is worth ₹6,000 + ₹60,000 = ₹66,000. The number of bad loans avoided is 70,000 × 0.006 = 420, and the gain is 420 × ₹66,000 = ₹277.2 lakh, or about ₹2.77 crore a year.
That number is a single point. The improvement of 0.6 points was estimated from a sample and carries uncertainty. Suppose the credible range is 0.3 to 0.9 points. Because the gain moves in proportion, ₹4.62 lakh for each 0.1 point, the range is ₹138.6 lakh to ₹415.8 lakh.
Predict first: the improvement is 0.6 points with an uncertainty of plus or minus 0.3 points, which is half of it. By what share will the rupee gain vary, about 10 per cent, about 50 per cent, or about 100 per cent? Decide before reading on.
About 50 per cent. The gain is directly proportional to the improvement, so an uncertainty of half the improvement is an uncertainty of half the gain, in this case ₹139 lakh either way. A tidy ₹2.77 crore without a range would claim far more precision than the evidence supports. The honest statement is "about ₹2.8 crore a year, plausibly between ₹1.4 and ₹4.2 crore," and it is the stronger statement, because a manager who sees only a point estimate and later gets half of it feels misled, while one who was told the range feels informed.
State the assumptions beside the number: same approval volume, same losses per bad, a stable economy, and the estimate of the improvement. If the result depends strongly on one assumption, say so and show the result when that assumption changes.
Charts that carry one message
The charts of Lesson 1.4 apply here with extra discipline. A chart in a presentation exists to make a point, so give each one a single message and say it in the title. "Bad rate falls from 20% in the riskiest band to 1.8% in the safest" tells the reader what to see. "Bad rate by band" makes them work it out.
Message · Good chart · AvoidCompare categories · Sorted bars, starting at zero · Pies with many slices, 3D effects
Show a trend · A line with the key points labelled · Dual axes that invite false links
Show a distribution · Histogram or box plot · A bare average
Show a relationship · Scatter plot, with a trend line only if justified · A line through a cloud
Show a range or uncertainty · Bars or points with intervals · A single point with no sign of doubt
Show parts of a whole · A few stacked bars, or a table · A pie with eight slices
Habits that make charts honest and clear: start bar charts at zero; label the lines or bars directly, not through a legend; use one colour to highlight the point and grey for the rest; put units in the axis label; and cite the source and the period under the chart. Remove everything that does not help the message, including gridlines, borders and decoration. Check the chart in the way a sceptic would: if the scale were changed, would the message look larger than it is?
Words that match the evidence
How you phrase a claim should match how strong it is.
Say the number with its unit and its comparison. "The bad rate is 3.4 per cent against 4.0 per cent today" is better than "the bad rate improves".
Match verbs to evidence. After a randomised test, "the offer increased conversion". After an observational analysis, "customers who received the offer converted more, though they may differ in other ways". Lesson 2.4 and Lesson 7.4 explain why the difference matters.
Give the limits in the same breath. "On 301 defaults and a later sample of 2,000 loans" is both honest and credible.
Separate the finding from the decision. The data says what happened; the decision depends on values and costs the analyst may not own.
Say what would change your mind. "If the out-of-time Gini falls below 0.34 I would stop the trial" shows that you have thought about being wrong.
When you are questioned, answer the question asked, say "I don't know, and here is how I'd find out" when that is true, and never defend a number you cannot reproduce. A pause and a clarifying question are signs of competence.
Your portfolio
A portfolio is a small collection of projects that shows, to someone who has never met you, what you can do and how you think. For a person entering the field it often counts for more than a list of courses, because it is evidence, not a claim.
What employers look for
Reviewers usually spend only a few minutes on a portfolio, so it must be easy to scan. They look for a real question, clean and reproducible work, sound judgement about validation and leakage, clear writing, and evidence that you can finish something. Quantity matters less than quality: three strong projects beat ten shallow ones.
A set of three
Project · Skills shown · Example from this courseA decision-focused analysis with a clear business question · SQL, cleaning, summaries, charts, a recommendation · Cohorts and funnels for a lender (Module 3)
A predictive model with honest validation · Features, a baseline, cross-validation, the right metric, explanation · The default model with leakage checks and SHAP (Modules 4 to 7)
An end-to-end project that reaches a decision · Target definition, scorecard or model, cutoff from economics, monitoring plan, documentation · The scorecard of Lesson 8.3
Choose public data, or data you have simulated and say so. Never publish confidential employer or customer data, even disguised. A project built on a well-known public dataset still stands out if the question is sharp and the validation is careful.
How to present each project
Give every project the same short structure, as a repository with a clear README and a one-page write-up.
1. The question and the decision in a sentence or two.
2. The data: source, size, period, and the definition of the target.
3. The approach: the baseline, the methods tried, how you validated, with the scores and their margins.
4. The result: the main finding in plain words and the value in real units.
5. Limits and next steps: what you would test next, and what could be wrong.
6. How to run it: a short list of commands, a file of package versions, and a fixed random seed, so that someone else can reproduce your numbers.
Keep the code tidy, with functions and tests as in Lesson 8.1, and keep the notebook clean enough to read from top to bottom. A short post or recorded walk-through explaining a project in your own words shows communication skill, which is what many teams struggle to find.
Resume lines that state results
A resume line should say what you did, how, and what came of it. A useful pattern is: action verb, method, number.
Weak: "Worked on a loan default model."
Stronger: "Built a credit scorecard on 5,000 loans using weight of evidence and logistic regression, reaching a Gini of 0.41 (0.38 out of time), and set a cutoff that raised simulated profit by 30 per cent."
Weak: "Used Python for data analysis."
Stronger: "Cleaned and merged three tables with pandas and SQL, found a 12-point drop in a conversion funnel at the document-upload step, and proposed a change tested in an A/B experiment."
Use only numbers you can defend in an interview, and expect to be asked about every word.
Interviews
Data science interviews vary, but most contain a mix of the same parts. Knowing the parts lets you prepare for each.
Stage · What is tested · How to prepareScreening call · Motivation, basic fit, a project summary · A 90-second story of your best project
SQL or coding exercise · Joins, grouping, windows, data cleaning with pandas · Practise Lessons 3.2 and 1.2 on fresh data, under time
Statistics and probability · Sampling, intervals, tests, base rates, A/B tests · Work the examples of Module 2 aloud
Machine learning concepts · Overfitting, regularisation, metrics, leakage, validation · Explain the ideas of Modules 4 to 6 in plain words
Case or product question · Structuring an open problem and choosing metrics · Practise the answer-first structure
Take-home project · A small end-to-end analysis · Use the portfolio template, and respect the time limit
Behavioural · Teamwork, conflict, ownership, mistakes · Prepare four stories and the lessons they taught
### Worked questions
A sample-size question. "We want to test whether a new email raises conversion from 5 per cent by 1 percentage point. How many customers do we need?" A quick, defensible answer uses the rule of thumb for a comparison of two proportions at the usual 5 per cent significance and 80 per cent power: the number in each group is about 16 × p × (1 − p) ÷ d², where d is the difference to detect. Here that is 16 × 0.05 × 0.95 ÷ 0.01² = 7,600 per group, or 15,200 in total. Then add the reasoning: if only 3,000 customers are available, you could test a larger effect, or run the test longer.
A base-rate question. "A fraud model catches 90 per cent of frauds and wrongly flags 2 per cent of genuine transactions. About 0.2 per cent of transactions are fraud. A transaction is flagged. How likely is it to be fraud?" Take a million transactions. There are 2,000 frauds and the model catches 1,800 of them. Among the 998,000 genuine ones, 2 per cent, 19,960, are flagged. So 1,800 of 21,760 flags are fraud: 8.3 per cent. The model is good, and most alerts are still false, because fraud is rare. This is the base-rate reasoning of Lesson 2.1, and a clear walk-through with round numbers is exactly what an interviewer hopes to see.
A concept question for a non-expert. "Explain overfitting to a manager." A strong answer is short and uses an analogy: "Overfitting is when a model memorises the past instead of learning the pattern, like a student who learns last year's exam answers and then fails a new paper. We catch it by testing on data the model has never seen. If the score is great on the old data and much worse on the new, it has memorised."
A project question. "Tell me about a project." Use a simple arc: the situation and the decision, your task, what you did and why, the result with a number, and what you learned or would change. For the scorecard: "A lender wanted to decide which personal loan applicants to approve. I built a compact scorecard from three variables using weight of evidence, checked it on a later sample, chose a cutoff of 550 from the profit and loss per loan, and wrote a monitoring plan. It would have earned about 30 per cent more than approving everyone. The part I'd redo is the sample, which had only 301 defaults, so I'd collect more before launch."
A judgement question. "Your model's Gini fell from 0.41 to 0.38 on new data. What do you do?" "First I'd ask how much data that rests on. With about 300 defaults the margin is about 0.07, so the fall is within noise. I'd keep watching the trend and check the inputs for drift, and I'd only retrain if the fall persisted or went below the agreed floor." This answer shows that you know about noise, monitoring and process.
Habits for the interview itself
Think aloud. The interviewer is judging your process, so say what you are assuming and why.
Clarify before solving. Ask what the metric is, what the data looks like, and what the decision is.
Start simple. Offer a baseline before a complex model.
Check your answer. Test a result against an extreme case or a rough estimate.
Own mistakes. Say "I made an error there, here is the correction." It is a good sign.
Prepare questions. Ask about the data, how models reach decisions, how success is measured and how monitoring is done. They show that you think like someone who ships.
Where to go from here
You now have a complete route through the subject: data and its cleaning; statistics and experiments; analytics methods; regression and its diagnostics; machine learning; validation; advanced methods; and delivery. Reading will not make it stick. Practice will. A sensible plan for the next three months is to do one project each month, finishing each with a write-up. Choose a domain you care about, because depth in a domain, such as lending, healthcare, retail or operations, is what turns a general skill into a valuable one. Strengthen SQL, because it is used daily. Practise explaining your work to a non-expert. And return to the earlier lessons when a project exposes a weakness; the course was built to be reread.
python
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
def yearlygain(apps, approve, oldbad, newbad, profitgood, lossbad):
"""Rupees gained when bad loans are replaced by good ones at constant volume."""
avoided = apps approve (oldbad - newbad)
return avoided, avoided (profitgood + lossbad)
central case and a range, from the uncertainty in the improvement
cases = {"Low": 0.034 + 0.003, "Central": 0.034, "High": 0.034 - 0.003} # bad rate under the new model
gain = {k: yearlygain(100000, 0.70, 0.040, v, 6000, 60000)[1] / 1e5 for k, v in cases.items()}
print({k: round(v, 1) for k, v in gain.items()}) # Rs lakh
fig, ax = plt.subplots(figsize=(6, 3.2))
bars = ax.bar(list(gain), list(gain.values()), color=["#9aa5b1", "#1f6feb", "#9aa5b1"])
ax.barlabel(bars, fmt="%.0f")
ax.setylabel("Rs lakh a year")
ax.settitle("The new scorecard adds about Rs 2.8 crore a year, plausibly Rs 1.4 to 4.2 crore")
for side in ("top", "right"):
ax.spines[side].setvisible(False)
fig.tightlayout()
fig.savefig("impact.png", dpi=150)
the rule of thumb for an A/B test, and the base-rate answer
def npergroup(p, d):
return 16 p (1 - p) / d 2
def alertprecision(rate, sensitivity, falsealarm):
return rate sensitivity / (rate sensitivity + (1 - rate) falsealarm)
print(round(npergroup(0.05, 0.01)), round(alertprecision(0.002, 0.90, 0.02), 3))
Common mistakes
Narrating the work in the order it was done. Start with the recommendation and add evidence as needed.
Reporting a model metric to a business audience. Translate into rupees, customers or hours.
A single number with no range. State the uncertainty, and the assumptions behind it.
Charts without a message. Put the finding in the title, start bars at zero and label directly.
Overclaiming from observational data. Match the verb to the evidence.
A portfolio of copied tutorials. Use your own question and careful validation, and say what is yours.
Showing confidential data. Use public or simulated data only.
Resume lines with no results. Say what you did, how, and what came of it, with numbers you can defend.
Memorised interview answers. Think aloud and reason from first principles, which is what is being tested.
Key takeaways
An analysis is done when someone acts on it. Start from the decision, and write for the most senior reader first.
Use an answer-first structure: recommendation, value, evidence, risks and limits, and the ask.
Translate model gains into rupees with a range. A 0.6-point fall in bad rate is worth ₹66,000 per avoided bad loan, ₹2.77 crore a year in the example, with a range of ₹1.39 to ₹4.16 crore.
Give each chart one message in its title, start bars at zero, label directly and show uncertainty. Match your wording to the strength of the evidence.
A portfolio of three well-finished projects, with a question, honest validation, a result in real units and instructions to reproduce it, is worth more than many shallow ones.
Interviews mix SQL, statistics, machine learning concepts, a case, perhaps a take-home and behavioural questions. Prepare by working examples aloud: 16 × p × (1 − p) ÷ d² gives 7,600 per group, and a 90 per cent test with a 2 per cent false alarm rate on 0.2 per cent fraud is right on only 8.3 per cent of alerts.
Keep learning by finishing one write-up a month in a domain you care about.
Check yourself
1. A new model lowers the bad rate by 0.4 points on 70,000 approved loans, with ₹6,000 profit per good loan and ₹60,000 loss per bad loan. What is the yearly gain?
Bad loans avoided: 70,000 × 0.004 = 280. Gain: 280 × ₹66,000 = ₹184.8 lakh.
2. Why is "₹2.77 crore a year" a weaker statement than "about ₹2.8 crore, plausibly ₹1.4 to ₹4.2 crore"?
The first claims more precision than the evidence supports. The second gives the range implied by the uncertainty in the estimate.
3. How many customers per group does a test need to detect a rise in conversion from 10 per cent by 2 points?
16 × 0.10 × 0.90 ÷ 0.02² = 3,600 per group.
4. Rewrite "Worked on customer churn" as a stronger resume line.
For example: "Built a churn model on 20,000 customers with gradient boosting and cross-validation, reaching an AUC of 0.78, and identified the 10 per cent of customers who account for 35 per cent of churn." Use only numbers you can defend.
5. In the fraud question, why is the alert precision only 8.3 per cent despite a good model?
Fraud is rare, 0.2 per cent. Even a 2 per cent false alarm rate on the huge genuine group produces far more false alerts, 19,960, than the model's 1,800 true catches.
6. Name two things to do at the start of an interview question before solving it.
Clarify the decision and the metric, and state your assumptions. Offer a simple baseline first.
Try this
Set aside one weekend to turn one of your course projects into a portfolio piece.
1. Choose the project, write its question and decision in two sentences, and name the audience.
2. Write the answer-first summary: recommendation, value, evidence, risks and the ask. Keep it under 150 words.
3. Compute the value in rupees, or the unit of your project, for a low, central and high case, and check your arithmetic with the impact simulator.
4. Make two charts, each with a title that states its finding. Check each against the habits listed above.
5. Create the repository: README with the six parts, a file of package versions, a fixed seed and a command that reproduces the main table.
6. Write three resume lines for the project, each with action, method and a number you can defend.
7. Rehearse a 90-second description of the project aloud, then answer the base-rate question and the sample-size question without notes. Ask a friend to interrupt with "why?" at every step.
8. Write a three-month plan: one project a month, the domain, and the lesson you will return to first.
A good result is a repository a stranger can run, a one-page summary a manager can act on and a spoken account you can give without notes. That completes the course. You can take a question from raw data to a validated model, a monitored service and a decision, and explain each step to the person who has to act on it.
Tags: Data Science Data storytelling Data science portfolio Interview preparation Communication