Skip to content

About

Bellabeat data analytics case study using Fitbit fitness tracker data.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

Β 

History

3 Commits

Folders and files

NameName
Last commit message
Last commit date
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

Bellabeat Case Study β€” How Can a Wellness Company Play It Smart?

Capstone project of the Google Data Analytics Professional Certificate (Case Study 2), following the six phases Ask, Prepare, Process, Analyze, Share and Act. Python handles data preparation, analysis and charts; reproducing the reference indicators in Tableau is planned as an independent check.

πŸ“„ Full report: report/analyze_report.md Β· PDF version Β· Decision log: docs/decision_log.md Β· Licences: LICENSE-content.md

Data source: the data were downloaded from the Kaggle data set FitBit Fitness Tracker Data (MΓΆbius). The Kaggle page credits the original authors (Furberg, Brinton, Keating and Ortiz) and links their Zenodo record. See Β§ 4 and Β§ 14.

Disclaimer: this is an independent educational project. It is not affiliated with, endorsed by or sponsored by Bellabeat, Fitbit, Google, Coursera or Kaggle.


1. Project overview

Bellabeat makes health-focused smart products for women (Leaf tracker, Time watch, Spring bottle, Bellabeat app, membership). This project analyses public usage data from non-Bellabeat fitness trackers to identify usage trends. It then turns these trends into high-level marketing recommendations for one Bellabeat product.

2. Business task

Analyse how consumers use non-Bellabeat smart devices, apply the insights to one Bellabeat product, and give high-level recommendations to guide Bellabeat's marketing strategy. The stakeholders are UrΕ‘ka SrΕ‘en (cofounder, Chief Creative Officer), Sando Mur (cofounder) and the executive team.

3. Questions addressed

  1. What are some trends in smart device usage?
  2. How could these trends apply to Bellabeat customers?
  3. How could these trends help influence Bellabeat marketing strategy?

4. Data

4.1 Data source

  • Data set used: FitBit Fitness Tracker Data on Kaggle (uploaded by MΓΆbius; licence shown on Kaggle: CC0). All files used in this project were downloaded from this page.
  • Original source: the Kaggle page credits Robert Furberg, Julia Brinton, Michael Keating and Alexa Ortiz and links their Zenodo record, Crowd-sourced Fitbit datasets 03.12.2016–05.12.2016 (RTI International, doi:10.5281/zenodo.53894). It is cited for attribution only; nothing was downloaded from Zenodo.
  • Scope used: export of 12 Apr – 12 May 2016, three files: dailyActivity_merged (940 rows, 33 Ids), sleepDay_merged (413 rows, 24 Ids), hourlySteps_merged (22,099 rows, 33 Ids).
  • Main limitations:
    • 33 Ids, where the source describes 30 consenting users; the Id column is an "export session ID".
    • One month of 2016 data, collected through a survey on Amazon Mechanical Turk.
    • No demographic variables, so the sex, gender or age of participants are unknown.
    • Sleep data cover only 24 Ids.
    • The sample is not representative of Bellabeat customers.

4.2 Reproducing the analysis

  1. Download the data set from the Kaggle page. Place the folder mturkfitbit_export_4.12.16-5.12.16 in data/raw/, which is ignored by Git.

  2. Check that your files match the ones used here (SHA-256 of the Kaggle download used for this project):

    File (Fitabase Data 4.12.16-5.12.16/) SHA-256
    dailyActivity_merged.csv 4b8f557eb29cd929cee7cc2d2a5e7704578bbffbc42c352e7b0c42bc4f3c6ac8
    sleepDay_merged.csv 2d4250ca47dfc74263f1b61fe4ba3e338c83faab0d352fb0fefa93464b170e53
    hourlySteps_merged.csv 83b9c1acc782b58a75a550f2f1d4ee9da7723e0ec26798b502768420cf6d7316
  3. Rebuild the prepared data. The script checks 18 expected counts and stops if the input differs. Add --overwrite if data/prepared/ already contains its quality files.

python scripts/process/bellabeat_process.py --raw-dir "data/raw/mturkfitbit_export_4.12.16-5.12.16/Fitabase Data 4.12.16-5.12.16" --out-dir data/prepared

Expected outputs, their hashes and column details: data/README.md. Analysis, charts and checks: Β§ 12.

4.3 Data availability

Raw and prepared row-level data are not hosted in this repository. They can be rebuilt from the Kaggle download with the steps above. Only aggregated results (tables and figures) and processing logs without individual Ids are published. The reason is that the redistribution rights of the downloaded files are not sufficiently established (see docs/licensing_note.md).

5. Tools and technologies

  • Python 3.14 with pandas, NumPy and matplotlib: cleaning, analysis and charts.
  • Windows PowerShell 5.1: read-only inventory and quality checks.
  • Tableau: planned independent reproduction of the reference indicators (not yet done).
  • Markdown: report and documentation.

6. Data preparation and quality controls

  • Raw files are never modified: their SHA-256 hashes are compared before and after the run.
  • Cleaning steps:
    • US date formats parsed explicitly;
    • Id kept as text;
    • 3 exact duplicates removed from sleepDay (413 β†’ 410 rows), the only rows deleted;
    • hourly totals joined to daily data in a 1-to-1 merge, with the row count checked (940 before and after).
  • Atypical values are flagged, never deleted: 72 days with 0 steps and 1,440 sedentary minutes (F1), 4 days with Calories = 0 (F3, kept as is), incomplete hourly coverage, and other flags F1–F9 and H1.
  • Controls run by the script: 18 expected counts, re-reading of the written files, unchanged totals. An independent verification script is also provided.

Details: report Β§ C, data/prepared/transformation_log_public.csv, data/prepared/quality_report_before_after.csv.

7. Analysis methodology

  • Descriptive statistics only: no significance tests, no causal claims.
  • Each result is reported two ways:
    • pooled (every Id-day weighs 1);
    • per-Id (each Id weighs 1).
  • Sensitivity analyses: without the 72 F1 days (S-F1) and without the 4 F3 days (S-F3). A change of 5 % or more, or any change of conclusion, is flagged.
  • External thresholds (10,000 steps, WHO 150 min/week, CDC 7 h of sleep) are used only as indicative reference points, with their sources checked.

8. Key findings

Descriptive observations for this sample; not causal and not generalisable to Bellabeat customers.

  • Activity data are recorded almost continuously; sleep data are not. All 33 Ids have activity data (median 31 of 31 days), but only 24 have any sleep record and only 12 have 20 nights or more. A sleep record exists for 43.6 % of days.
  • Most classified time is sedentary or light.
    • 81.3 % of classified minutes are sedentary, 15.8 % light, and 2.8 % fairly or very active.
    • Sedentary minutes depend on whether sleep was recorded, so they are not directly comparable across all days.
  • Daily steps vary widely.
    • Median: 7,406 steps per day.
    • 32.2 % of days reach 10,000 steps.
    • Mean daily steps range from 7,638 (all days) to 8,271 (without the F1 days).
  • Clear activity rhythm.
    • Weekday peak at 17:00–19:00 (640 steps/h at 18:00).
    • Weekend peak at 12:00–14:00 (639 steps/h at 13:00).
    • Sunday is the least active day in every specification tested.
  • Sleep. Among Ids with sleep records, the median time asleep is about 7 h 13, and 55.9 % of nights reach 7 hours.
  • Steps and calories. Within each Id, days with more steps go with higher estimated calories (median r = 0.87 across 32 Ids). This is an association, not a causal link.

9. Visualizations

Distribution of daily steps Mean daily steps by weekday
Fig. 1 β€” Daily steps are widely spread; about one day in three reaches 10,000. Fig. 2 β€” Sunday is the least active day in every specification.
Hourly step profile Composition of classified minutes
Fig. 3 β€” Weekday peak in the early evening; weekend peak around midday. Fig. 4 β€” About four in five classified minutes are sedentary.
Data coverage by feature Sleep duration
Fig. 5 β€” Activity is recorded for every Id; sleep, heart rate and manual logs far less. Fig. 6 β€” Just over half of recorded nights reach 7 hours.

10. Three marketing recommendations

These recommendations are hypotheses derived from a small 2016 sample of non-Bellabeat users. They are not validated with Bellabeat customers and should be tested on Bellabeat's own data. Product focus: Leaf or Time, with the Bellabeat app. The final product choice is still open.

  1. Make consistent sleep tracking a central promise and an onboarding goal. Activity is recorded for all Ids but sleep only partially (24 of 33 Ids, 12 regularly). The data show that sleep is recorded less often, not why.
  2. Replace a single 10,000-step message with personalised, progressive activity goals. Activity levels differ widely between Ids, and only 32.2 % of days reach 10,000 steps, which is a popular target rather than an official guideline.
  3. Time reminders and content to the observed activity rhythm, then A/B test them. The weekday evening peak, the weekend midday peak and the low-activity Sundays are stable patterns. Whether users respond better to messages at those times is untested.

Full reasoning, evidence and limits: report Β§ 10.

11. Project structure

bellabeat-case-study/
β”œβ”€β”€ README.md
β”œβ”€β”€ requirements.txt
β”œβ”€β”€ .gitignore Β· .gitattributes
β”œβ”€β”€ LICENSE (MIT, code) Β· LICENSE-content.md (licence scope, CC BY 4.0 content, third-party exclusions)
β”œβ”€β”€ report/analyze_report.md          # full report: six case-study deliverables
β”œβ”€β”€ data/
β”‚   β”œβ”€β”€ README.md                     # sources, licences, how to obtain raw data
β”‚   └── prepared/                     # public transformation log + quality report (cleaned CSVs rebuilt locally, not published)
β”œβ”€β”€ scripts/
β”‚   β”œβ”€β”€ process/                      # cleaning (bellabeat_process.py) and verification
β”‚   β”œβ”€β”€ analysis/                     # descriptive analysis -> results/tables
β”‚   β”œβ”€β”€ visualization/                # charts -> results/figures
β”‚   β”œβ”€β”€ checks/                       # read-only inventory and quality checks
β”‚   └── publication/                  # builds the *_public files (no raw rows, no individual Ids)
β”œβ”€β”€ results/
β”‚   β”œβ”€β”€ tables/                       # aggregated tables (CSV), analysis_results.json, per-Id tables as *_public summaries
β”‚   └── figures/                      # six PNG charts
β”œβ”€β”€ docs/
β”‚   β”œβ”€β”€ decision_log.md               # validated decisions and script provenance
β”‚   β”œβ”€β”€ licensing_note.md             # rights by content type (licence not yet chosen)
β”‚   └── raw_file_inventory_public.csv # inventory of the 29 raw files (metadata only)
└── tests/README.md                   # how to run the verification checks

12. Reproducibility instructions

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

The commands below are run from the repository root.

1. Rebuild the prepared data from the Kaggle download: see Β§ 4.2. To check the input without writing anything, add --dry-run instead of --out-dir.

2. Analysis and charts (from the rebuilt prepared data):

python scripts/analysis/bellabeat_analyze.py --prepared-dir data/prepared --out-dir results/tables
python scripts/visualization/bellabeat_share_charts.py --prepared-dir data/prepared --out-dir results/figures

The analysis script also writes per-Id tables (a1_usage_per_id.csv, a5_sleep_per_id.csv) and the cleaning script writes transformation_log.csv with Ids. These files stay local (ignored by Git); their public versions are rebuilt with:

python scripts/publication/make_public_versions.py

3. Checks. verify_outputs.py requires the raw data and the run_metadata.json produced by step 1; counter_check.py requires the rebuilt prepared data. See tests/README.md.

python scripts/process/verify_outputs.py
python scripts/checks/counter_check.py

All scripts were run on Windows with Python 3.14.7 and reproduced the original outputs byte for byte (expected hashes of the prepared files in data/README.md). Other platforms and versions have not been tested.

13. Limitations and ethical considerations

  • Small, self-selected sample (33 Ids, recruited through Amazon Mechanical Turk); one month in 2016; Fitbit users only.
  • No demographic data. The project makes no assumption about the sex, gender or age of participants, and does not equate them with Bellabeat's customers.
  • Calories, intensity minutes and sleep are device estimates from undocumented algorithms. Sedentary minutes depend on sleep tracking.
  • Participants are identified only by numeric session Ids, with no direct identifiers. Neither raw nor prepared row-level data are redistributed in this repository; published tables are aggregated, and the few Ids mentioned in processing logs are replaced by neutral references.
  • Results are descriptive associations; the recommendations are hypotheses to test.

14. References and data attribution

  • MΓΆbius. FitBit Fitness Tracker Data [Data set]. Kaggle. https://www.kaggle.com/datasets/arashnic/fitbit (licence shown on Kaggle: CC0). Source of the files used in this project.
  • Furberg R., Brinton J., Keating M., Ortiz A. (2016). Crowd-sourced Fitbit datasets 03.12.2016–05.12.2016 [Data set]. Zenodo. https://doi.org/10.5281/zenodo.53894 (licence shown on Zenodo: CC BY 4.0). Original source credited by the Kaggle page.
  • Google / Coursera. Case Study 2: How can a wellness company play it smart? Google Data Analytics Professional Certificate (case brief; not redistributed here).
  • Tudor-Locke C., Bassett D.R. (2004). How many steps/day are enough? Sports Medicine, 34(1), 1–8.
  • World Health Organization (2020). WHO guidelines on physical activity and sedentary behaviour.
  • Centers for Disease Control and Prevention. About Sleep. https://www.cdc.gov/sleep/about/

Data attribution

The data were downloaded from the Kaggle data set FitBit Fitness Tracker Data (MΓΆbius). Kaggle shows the licence CC0, and use of the site is subject to Kaggle's terms of use. The Kaggle page credits the original authors, Robert Furberg, Julia Brinton, Michael Keating and Alexa Ortiz (RTI International), and links their Zenodo record (https://doi.org/10.5281/zenodo.53894), which shows the licence CC BY 4.0. The Kaggle files have not been compared with the Zenodo archives, so this project does not claim that the Zenodo licence applies to the Kaggle copy. The original authors are credited in any case.

Changes made in this project: duplicate removal, date formatting, derived columns and quality flags, aggregation into statistics and charts. Raw and prepared row-level files are not distributed here (see Β§ 4.3).

Licences

  • Code (scripts/ and the hand-written LaTeX sources in report/latex/): MIT License, Copyright (c) 2026 LUMU BAMANA MANASSE.
  • Original report text, figures, aggregated tables and documentation: CC BY 4.0, Copyright (c) 2026 LUMU BAMANA MANASSE (author of the report; GitHub pseudonym MediJr).
  • Third-party material is excluded from both licences: the data set, the case-brief scenario and questions, quotations, trademarks, the fonts embedded in the PDF and the software dependencies.

Exact scope: LICENSE-content.md. Background: docs/licensing_note.md.

Disclaimer

This is an independent educational project for the Google Data Analytics Professional Certificate. It is not affiliated with, endorsed by or sponsored by Bellabeat, Fitbit, Google, Coursera or Kaggle. Their names are trademarks of their respective owners.

About

Bellabeat data analytics case study using Fitbit fitness tracker data.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages