Data Analytics Portfolio

Three projects, seven sections each

Robert N. Dove · data cleaning, regression, dashboards and measurement integrity

Each project below is written in the seven sections a hiring manager actually scans: the question, where the data came from, what it was analysed with, what was done to it, what came out, and where to see it. Every figure is quoted from the analysis file itself rather than from a summary. Two further projects — a marketing attribution audit and a conversion integrity review, both on live client revenue data — are written up on the projects page.

Key insights & results

1. Termination and retirement were tracked separately, against separate targets. Company targets of 25% and 10% respectively were held as explicit thresholds rather than folded into a single attrition number, which surfaced exactly which departments breached which target. Collapsing the two would have hidden the distinction the Chief People Officer actually needed to act on.

2. Handling blank fields correctly was the whole analysis. Attrition date is empty for active staff, so a naive filter on that column would have silently dropped every current employee from the counts — producing a dashboard that reported 100% attrition and looked entirely plausible. The blank is structural: it means "still here", not "missing data".

3. Five characteristics were tested against attrition category. Age, Monthly Income, Years at Company, Distance From Home and Total Working Years were each analysed against the Terminated / Retired / No Attrition classification, so the dashboard answers who is leaving as well as how many.

Link. View the live dashboard in Google Sheets · all nine tabs are protected, so the calculation layer cannot be altered by a reader · a presentation recording is available on request.

Key insights & results

1. The headline finding reverses the obvious reading of the data. Across the whole fleet, the displacement–MPG R² sits near 0.60 through the 1990s, 2000s and 2010s and then falls to 0.4451 in the 2020s — which looks like engine size losing its meaning as the fleet electrifies. Comparing like with like reverses it. Among naturally aspirated, non-electrified vehicles the relationship strengthens in every successive decade: 0.6256 → 0.6809 → 0.7327 → 0.7862. The apparent modern collapse is an artefact of the fleet becoming less homogeneous, not of displacement mattering less.

2. A routine cleaning rule would have destroyed the dataset. A blanket “drop every row containing a blank” leaves zero of 41,837 rows, because several columns use blanks structurally to mean a feature is absent. The 1,615 missing displacement values are not random either: 1,571 are electric, 42 fuel-cell, 2 eFCV — displacement is simply not a combustion measure for them.

3. One row changed the model. A 2016 Mitsubishi i-MiEV carried a recorded displacement of 0.0 L. It is electric, so that zero is a category error rather than a measurement. Excluding it is the difference between the corrected model and the naive one.

4. The two most tempting predictors were rejected on principle. City and highway MPG correlate almost perfectly with combined MPG — because combined MPG is constructed from them. Using them would be target leakage. Displacement was chosen precisely because it is an independent physical characteristic of the vehicle.

What it means for a buyer. Fitted on the 40,221 vehicles carrying a positive displacement, the model explains 51.9% of the variance in combined MPG (R² 0.5192). Fitted equation: combined MPG = 31.0989 − 3.0879 × displacement, so roughly 3.09 MPG lost per additional litre. But prediction error is 3.95 MPG — 18.8% of the sample average — and the residuals are uneven, so this is a screening rule, never a substitute for the published figure. A smaller engine remains a strong efficiency signal, provided you compare like with like.

Method note: residual diagnostics were reported honestly rather than skipped — the mean residual is essentially zero as expected, but 80.14% of residuals fall within one standard deviation against a textbook 68%, and the spread widens at both ends of the displacement range. That is nonlinearity and heteroscedasticity, and it is the reason the write-up recommends displacement as a screen rather than a predictor.

Link. The full 11-page manager-facing report and the underlying workbook are available on request — the source dataset is the public EPA fuel economy file, so the analysis is reproducible from the public EPA fuel economy dataset and epa_fuel_economy.csv and the method described above.

Key insights & results

1. The first version of the report answered a question nobody asked. It led with a record count per species. Because the dataset holds exactly 50 rows for each of the three species, that visual reported the same number three times — it was measuring the sampling design, not the flowers. It was replaced with aggregated measurements, which is the comparison a reader actually came for.

2. Petal measurements separate the species; sepal measurements do not. Setosa is consistently the smallest and Virginica the largest, but the gap is far cleaner on petal length and width than on sepal width, where the three species overlap heavily. A reader filtering to a single species sees that immediately in the scatter chart.

3. The theme was authored, not chosen. The report ships a custom theme file defining eight data colours plus background, foreground, good/bad/neutral, table accent and text classes — so every visual encodes the same species in the same colour, and the report stays readable when a slicer changes what is on screen.

Honest scope. This is a visualisation exercise on a small, clean, famous dataset, and it is presented as one. The aggregation is done through Power BI’s field wells; the report does not use DAX measures or Power Query transformations, and no claim is made that it does. What it demonstrates is report construction, cross-filtering behaviour and deliberate visual encoding.

Link. The Power BI report is a .pbix file, which is not viewable in a browser without a Power BI licence, so it is offered as a file on request rather than as a dead link. The dataset itself is public — the Iris dataset at UCI — so every figure above can be reproduced. The authored theme (Dove_Robert_IrisTheme.json) travels with it.