# Session 1 Handout: How Predictions Get Made

<p class="meta">FAST-05 Applied Data Science and ML for Business Users · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

You already work with predictions: lead scores in the CRM, reorder suggestions in the ERP, fraud or quality flags, staffing forecasts. You don't need to build models to use them well. You need to know what goes in, what comes out, and who decides. This session gives you that vocabulary and a one-page prediction map.

## Key concepts

### 1. Rules vs. learned patterns

A **rule** is written by a person: "Flag claims over $10,000." "Reorder below 200 units." It's transparent and does exactly what it says.

A **model** learns patterns from many past examples: "Claims filed soon after a policy starts, at night, with no police report have tended to be fraud." It can find patterns people miss, but it's only as good as the history it learned from, and it can be confidently wrong when the world changes.

### 2. Training data, features, labels

| Term | Plain meaning | Clinic no-show example |
| --- | --- | --- |
| Training data | Past examples the model learns from | 2023-2024 appointments |
| Features | Inputs known at prediction time | Lead time, weekday, reminder sent, prior misses |
| Label | The past outcome to learn | Attended or no-show |
| Prediction | The model's output for a new case | Risk score 0.62 for next Tuesday 9 AM |

Supervised learning = learning from examples where the label is known.

### 3. Confidence scores

Most business models output a **score** (often 0 to 1). A score **ranks** cases by how much they resemble past positives. A score of 0.55 is not "55% likely" unless the model has been specifically calibrated, and it's never a decision. Someone chooses a **threshold** ("call patients above 0.6") and someone **acts**. Those are business choices.

### 4. Data rules for analysis tools

Copilot in Excel, Power BI Copilot, ChatGPT data analysis, and Gemini in Sheets make it easy to upload whole files. Before you do:

- use only approved tools with work accounts;
- remove columns you don't need (names, IDs, contact info);
- never upload personal information (G.L. c. 93H, 201 CMR 17.00) or PHI unless the tool is approved for it under a business associate agreement;
- "de-identified" has a specific legal meaning under HIPAA. Removing names alone is not enough.

## Worked example: prediction map

| Field | Charles River Family Dental no-show score |
| --- | --- |
| Predicts | Whether a booked appointment will be missed |
| Inputs | Lead time, weekday, reminder sent, prior misses |
| Learned from | 2023-2024 appointment outcomes |
| Output | Score 0-1 shown on the schedule |
| Who acts | Front desk calls appointments above 0.6 the day before |
| Wrongly says yes | An extra call to a reliable patient (minor annoyance) |
| Wrongly says no | Empty chair: lost hygienist hour (~$150, fictional) |
| Data notes | IDs only in analysis; no diagnosis or health data |

## Checklist: understanding a prediction

- [ ] I can say what it predicts and from what inputs.
- [ ] I know what past outcome it learned from, and how old that data is.
- [ ] I know who acts and what they do.
- [ ] I can state the cost of each error direction.
- [ ] I know what data I may and may not put into analysis tools.

## Do and don't

| Do | Don't |
| --- | --- |
| Ask "who acts on this score?" | Treat a score as a decision |
| Describe features in plain words | Assume the model "understands" your business |
| Check how old the training data is | Assume past patterns still hold |
| Strip unneeded columns before analysis | Upload whole customer files "for context" |
| Treat a score as a ranking | Read 0.55 as "55% chance" without checking |

## Try this at work

1. List every score, flag, or forecast you see in your systems.
2. Pick one and complete a prediction map.
3. Ask its owner one question you couldn't answer.

## Glossary

| Term | Meaning |
| --- | --- |
| Model | A pattern learned from past data that makes predictions |
| Supervised learning | Learning from examples with known outcomes |
| Feature | An input to the model |
| Label | The known outcome in training data |
| Score | Model output ranking how much a case resembles past positives |
| Threshold | The score above which the business takes action |

## Key takeaways

- Models learn patterns from past examples; rules are written by people.
- Scores rank; thresholds and people decide.
- Map every prediction: inputs, output, who acts, cost of errors.