Understanding what
moves occupancy.
Working with Avanath Capital Management through UCI’s MSBA program, I explored changes in affordable-housing occupancy and how the findings could inform leasing and marketing decisions.
- records brought together
- 7,000+
- internal and external sources
- 6
- monthly data coverage
- 2016–2024
- models compared
- 4
Tools used
From data preparation to client reporting.
AlteryxData preparation
PythonAnalysis & modeling
TableauVisual exploration
Power BIDashboard reporting
My role — Student Data Analyst. I contributed across all four questions, from cleaning data in Alteryx and building models in Python to visualizing findings and discussing them with the client. The results below reflect our team’s work.
Occupancy measures the share of available units that are occupied. We combined property history with real estate market data, unemployment indicators, and area median income to examine both portfolio-wide patterns and individual properties.
When do vacancies rise,
and how could teams respond?
Looking beyond the occupancy rate helped us identify conditions worth watching.
I compared nine years of monthly occupancy history with unemployment and mortgage rates. Some occupancy declines appeared alongside, or after, changes in those indicators. The charts gave us a starting point for investigating risk, although they did not establish what caused the declines.
Our recommendation was to use changes in external conditions as a prompt for a closer property review. That could help leasing teams investigate a problem earlier, before deciding whether a campaign or another operational response was appropriate.
Look for changes
Compare monthly occupancy patterns with market and economic indicators.
Check local context
Ask whether leasing activity or property conditions help explain the change.
Decide on a response
Use the review to guide outreach, leasing, or further investigation.
Based on the time-series analysis in presentation slides 11–13. The relationships were exploratory and were not statistically tested.
How well could we
predict occupancy?
XGBoost performed best in our reported comparison, with an R² of 0.9707 and an RMSE of 2.15 percentage points.
I worked on predictive modeling using occupancy history and selected market and economic variables. We compared four approaches: linear regression, a generalized additive model (GAM), random forest, and XGBoost. XGBoost had the highest model fit and the lowest RMSE in the original, untransformed comparison.
| Model | R² | RMSE (percentage points) |
|---|---|---|
| XGBoost | 0.9707 | 2.15 |
| Random forest | 0.9675 | 2.27 |
| GAM | 0.9586 | 2.55 |
| Linear regression | 0.9530 | 2.72 |
These results gave us a basis for discussing how forecasts might support planning and help teams spot properties that need attention.
How to read the result: R² measures model fit, so 0.9707 does not mean “97% prediction accuracy.” The deck does not describe a test on a later, unseen period or a comparison with using last month’s occupancy as the forecast. These figures summarize the reported model comparison; they do not establish future forecast accuracy. Source: appendix, slide 38.
Which factors matter,
and do they vary by property?
Last month’s occupancy was the strongest portfolio-level predictor. Individual properties told a more varied story.
I worked on the analysis at both levels: looking for patterns across the portfolio, then examining how individual properties behaved. Prior-month occupancy ranked highest across the models. The contribution of other variables, including rents, construction activity, cap rates, and unemployment, varied by model.
One comparison made the difference clear. Two property models had reported R² values of 0.94 and 0.43. The same general approach explained much more of the variation in one property’s occupancy than the other’s.
Property A
More of the observed variation was explained.
Property B
Less of the observed variation was explained.
Adapted from slide 23. Two anonymized examples; identifying property details are omitted.
That finding shaped our recommendation to look at each property’s context before choosing an intervention. A lower R² can also reflect limited variation in occupancy; it does not, on its own, mean that absolute prediction errors are large.
How could the findings guide
pricing and marketing?
We turned the analysis into four recommendations for forecasting, monitoring, and deciding where to focus effort.
I helped connect the findings to questions the client could act on: which properties deserved a closer look, what might explain a change, and whether a leasing or marketing response made sense.
Support portfolio planning
Bring occupancy forecasts into revenue and cash-flow discussions so teams can consider potential changes earlier.
Flag properties for review
Monitor forecasted declines and external indicators to help teams decide where to investigate.
Tailor leasing and marketing
Review campaigns and pricing in the context of the property and its local market, using the analysis as one input.
Match the response to the property
Use property-specific findings alongside the team’s operational knowledge to choose an appropriate next step.
Adapted from slide 26. Recommendations were presented to the client. Their impact after implementation was not measured in this project.
This project helped me connect model results to a business conversation: what the analysis shows, where it needs more context, and what a team could reasonably do next.