📈 Causal Impact Analysis Dashboard (Streamlit)

Analyze the effect of an intervention on a time series using covariates with Google's CausalImpact (TensorFlow port) — all through a simple Streamlit UI.

---

TL;DR (Quickstart)

# 1) Install deps (ideally in a fresh venv)
pip install -r requirements.txt

# 2) Launch the app
streamlit run app.py

Then open the local URL printed in your terminal (usually http://localhost:8501).

---

What this app does

• Lets you upload a CSV or Excel file containing a date column, a response variable, and optional control variables.
• Helps you pick an intervention date and pre/post periods.
• Runs CausalImpact to estimate the impact and shows:
  • A short summary
  • A more detailed report
  • A visual plot of actual vs. counterfactual

Under the hood it uses tfcausalimpact via from causalimpact import CausalImpact, with Streamlit and Matplotlib for the interface and charts.

---

Data format & tips

Accepted files

• CSV (.csv)
• Excel (.xlsx)

Required columns

• At least one column with "date" in its header (case-insensitive), e.g. Date, DATE, event_date, etc.
• A response (dependent) variable column (e.g. y, sales, visitors).
• Optional control variables (explanatory covariates).

Date values

• Best supplied as numeric YYYYMMDD (e.g. 20240131).
  The app will convert these to real dates for you.

Special notes for Excel

• If your Excel file has top matter (titles, notes, logos), that's fine — the app skips early rows and drops all-empty columns automatically.
• Make sure the first row after any top matter contains actual headers for each column.

Minimal example (CSV)

date,y,x1,x2
20240101,100,23,0.7
20240102,98,20,0.8
20240103,110,21,0.6
...

---

How to use the app

1. Start the app: streamlit run app.py
2. Upload your CSV/Excel file.
3. Pick the date column when prompted.
4. Choose:
   • the response column (what you’re measuring the impact on),
   • any control columns (covariates).
5. Set the intervention date (the date the change/treatment happened).
6. Confirm pre/post periods (defaults are pre = start → day before intervention; post = from intervention → last date).
7. Click Run Causal Impact Analysis.
8. Read the summary, detailed report, and inspect the plot.

Tip: Pre/post periods must exist in your data. Controls should be available across the full pre & post windows.

---

Interpreting results (at a glance)

• Pointwise & cumulative effects: How the response diverged from the counterfactual.
• Posterior intervals: Uncertainty bands for effects.
• Summary/report: Text interpretation from the model.

If your controls don’t explain the response well in the pre-period, the counterfactual may be weak. Try better covariates or a longer pre-period.

---

Local development

Prereqs

• Python 3.9+ recommended
• pip

Setup

python -m venv .venv        # optional but recommended
# Windows: .venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate

pip install -r requirements.txt
streamlit run app.py

The app will print a local URL (typically http://localhost:8501).

---

Run with Docker (local)

Build the image:

docker build -t causal-impact-app .

Run the container:

docker run -p 8501:8501 causal-impact-app

Then open http://localhost:8501 in your browser.

If deploying to platforms that expect port 8080 (e.g., Cloud Run), ensure the container is configured to serve on 0.0.0.0:8080.

---

Deploy to Google Cloud Run (one-liner)

Make sure you’re authenticated (gcloud auth login), have a project selected, and Artifact Registry & Cloud Run are enabled. Then run:

gcloud builds submit --tag europe-west2-docker.pkg.dev/PROJECT_ID/streamlit-repo/streamlit-app && gcloud run deploy streamlit-app --image europe-west2-docker.pkg.dev/PROJECT_ID/streamlit-repo/streamlit-app --platform managed --region europe-west2 --allow-unauthenticated --memory=2Gi --max-instances=1

Replace PROJECT_ID with your actual Google Cloud project ID.

Notes

• The single-line command above works in Windows CMD, PowerShell, and macOS/Linux shells.
• First run may take several minutes while the image builds and pushes to Artifact Registry.

---

Troubleshooting

• “No column with 'date' found”
  Ensure at least one header contains the word “date” (any case).

• Dates aren’t recognized
  Provide numeric YYYYMMDD values (or clean your date column before upload).

• Empty or malformed data
  Check that your first actual row after any prelude rows contains real headers and that columns aren’t entirely empty.

• Plot doesn’t render
  Try a smaller date range, ensure data types are numeric where expected, and that pre/post windows exist in your index.

• Weird results
  Try improving control variables, lengthening the pre-period, or removing noisy controls.

---

Tech stack

• Streamlit UI
• tfcausalimpact for Bayesian structural time-series
• pandas / numpy for data handling
• matplotlib for charts

---

Privacy & data handling

Uploaded files are processed in-memory by the app for the duration of your session; the app doesn’t write your data to disk unless you modify the code to do so.

---

License

Add your preferred license here (e.g., MIT).

---

Credits

• Google’s original CausalImpact methodology
• The community tfcausalimpact implementation