# Session 4 Handout: Forecasts and AI-Assisted Analysis

<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

Forecasts drive staffing, inventory, and cash. A point forecast hides how uncertain it is, and models don't know what your sales team heard yesterday. This session covers reading forecast ranges, measuring error, deciding when a person should override, and using AI analysis tools without trusting them blindly.

## Key concepts

### 1. Forecast ranges

A forecast like "1,330 (80% range 1,120-1,540)" means: the middle estimate is 1,330, and in roughly 8 of 10 similar months actual demand should fall in the range. Plan around the range: buffer stock, overtime options, or supplier flexibility toward the high end when stockouts are costly.

### 2. Measuring forecast error

| Measure | How | Use |
| --- | --- | --- |
| Absolute error | \|Actual − Forecast\| | Per period |
| MAE | Average of absolute errors | Typical miss, in units |
| % error | Absolute error / Actual | Compare items of different sizes |
| Outside range | Actual < Low or > High | Signal to investigate |

Measure separately before and after known changes; a single MAE can hide a shift.

### 3. When a person should override

Override when you have **specific, new information the model doesn't**: a confirmed contract, a plant shutdown, a promotion, a supplier failure. Don't override on a hunch. Every override gets a log: original, new value, reason, source, approver, review date. Then compare overridden and original forecasts against actuals.

### 4. AI-assisted analysis

| Tool | Useful for |
| --- | --- |
| Excel Forecast Sheet | Quick statistical forecast with a range (needs enough history) |
| Copilot in Excel | Plain-language questions, formula columns, highlights |
| Power BI | Copilot summaries; anomaly detection; key influencers |
| ChatGPT data analysis / Gemini in Sheets | Profiling, calculations, charts with shown working |

Rules: name the table and columns; ask a specific question; ask it to show the formula or calculation; **confirm one finding by hand**. Feature availability varies by license and changed during 2026; confirm in your tenant.

## Worked example: Bayside BP-2207

| Month | Forecast | Range | Actual | Abs error | Outside? |
| --- | --- | --- | --- | --- | --- |
| Jan | 1,200 | 1,050-1,350 | 1,180 | 20 | No |
| Feb | 1,220 | 1,060-1,380 | 1,250 | 30 | No |
| Mar | 1,250 | 1,080-1,420 | 1,310 | 60 | No |
| Apr | 1,270 | 1,090-1,450 | 1,240 | 30 | No |
| May | 1,290 | 1,100-1,480 | 1,620 | 330 | Above |
| Jun | 1,310 | 1,110-1,510 | 1,700 | 390 | Above |

MAE Jan-Apr = 35; May-Jun = 360. **Manual check** of the AI's "average actual = 1,383": 8,300 / 6 = 1,383.3 ✓.

**Override:** Sales confirmed Granite State Hydraulics' new program (+~350/month from May). July forecast 1,330 → override **1,680**. Logged: source = sales confirmation; approver = operations manager; review = August actuals.

## Checklist: forecast review

- [ ] Range read, not just the point.
- [ ] Errors computed; outside-range months flagged.
- [ ] Error measured before/after known changes.
- [ ] One AI finding confirmed by hand.
- [ ] Override only with new information; logged.
- [ ] Override reviewed against actuals.

## Do and don't

| Do | Don't |
| --- | --- |
| Plan to the range | Treat the point as certain |
| Investigate out-of-range actuals | Delete "odd" months |
| Override with logged evidence | Override on hunches |
| Ask the tool to show its working | Paste AI numbers into plans unchecked |
| Note when history is too short | Trust a forecast from six data points |

## Try this at work

1. Pull six to twelve months of forecast vs. actuals for one item.
2. Compute errors and flag outside-range months.
3. Ask an AI tool for the MAE, then check it.
4. Start an override log.

## Glossary

| Term | Meaning |
| --- | --- |
| Point forecast | Single middle estimate |
| Forecast range (interval) | Span likely to contain the actual |
| MAE | Mean absolute error |
| Override | Human replacement of a forecast value with a logged reason |
| Forecast Sheet | Excel feature producing a forecast with confidence bounds |

## Key takeaways

- Plan around ranges, not points.
- Override with new information, and log it.
- Confirm one AI finding by hand every time.