Analyzing Trends in Office Attendance In Major US Markets

Team ACH: Anh, Caden, and Heidi!

Outline



  1. Motivation and Research Question
  2. Data Sourcing
  3. Analysis and Results
  4. Future Directions

Motivation


The Challenge:

  • “Inform Savills of notable trends or microtrends in the commercial real estate market that could be used to advise clients on where, when, whether and how to locate their offices.”

Our Interest:

  • Understanding trends in why people show up to the office, so that clients can optimize their office size.

‘Office Occupancy’ As A Response


  • Data is collected by Kastle


  • Tells us proportion of people who attend the office in comparison to March 1, 2020


  • We want to be able to understand the trend of occupancy, and eventually predict occupancy for the future

Major Markets

Figure 1: Major Markets Recorded by Kastle

Office Occupancy

Figure 2: Seasonal Time Series of Occupancy

Office Occupancy: Deseasonalized

Figure 3: Trend Component of Occupancy (After LOESS Decomposition)

Factors That Affect Office Attendance


  1. Extreme Weather Events
  2. Unemployment Rate
  3. Traffic Congestion
  4. Covid Cases
  5. State Political Affiliation

Example Factor: Political Affiliation

Figure 4: State Political Affiliation

Office Occupancy: Deseasonalized

Figure 5: Trend Component of Occupancy (After LOESS Decomposition)

Our Research Question



What variables explain the increasing trend of post-COVID office occupancy across major markets?

Analysis: Methods


  • Bayesian Model


  • Regularization of the predictors


  • Hierarchical multivariate normal regression

Analysis: Model


  • For major markets \(j \in [1, 10]:\) \[ \begin{aligned} \sigma^{(j)} &\sim \text{Exponential}(0.1) && \text{Unexplained variation (noise) in this market} \\ \beta_0^{(j)} &\sim \mathcal{N}(0, 1) && \text{Intercept (baseline occupancy for this market)} \\ \tau^{(j)} &\sim \mathcal{C}^{+}(0, 1) && \text{How complex the model is "allowed to be" in this market} \\ \end{aligned} \]

  • For explanatory variables \(i \in [1, 5]:\) \[ \begin{aligned} \lambda_i^{(j)} &\sim \mathcal{C}^{+}(0, 1) && \text{How much the variable "matters" in this market}\\ \beta_i^{(j)} &\sim \mathcal{N}\big(0, (\lambda_i^{(j)})^{2} \cdot (\tau^{(j)})^{2}\big) && \text{Overall effect of the variable in this market} \\ \end{aligned} \]

  • Model: \[ \begin{aligned} \mu_t^{(j)} &= \beta_0^{(j)} + \sum_{i=1}^5 \beta_i^{(j)} x_{i,t} &&\text{Predicted occupancy at time } t \text{ for this market} \\ M_t^{(j)} &\sim \mathcal{N}(\mu_t^{(j)}, \sigma^{(j)}) && \text{Observed occupancy rate (data likelihood)} \end{aligned} \]

Results: One Covariate for All Markets

Figure 6: Relative importance of political affiliation as predictor of office occupancy

Results: All Covariates for One Market

Figure 7: Chicago Coefficient Estimates

Results: All Markets

Figure 8: All Market Coefficient Estimates

Model Verification

Figure 9: Posterior Prediction

Conclusions

  • Different variables have unique impacts on major markets

  • The model is easily extendable to add more covariates

  • Our model gives useful results in two main ways:

    • For Savills: Better understand which properties to advertise to their clients
    • For the Client: Understand which market is best suited for them based on external factors

  • We were surprised by the relationship with traffic congestion

Future Directions


  1. We were very limited by data collection


  1. Want to do prediction for a client’s future office occupancy


  1. Increasing interpretability for clients

Credit

Thanks!