In early 2020, Covid-19 swept the world. To contain the virus and “flatten the curve” a number of regions implemented work-from-home policies (“Covid-WFH”) designed to reduce the spread through person-to-person contact.
Although Covid-WFH policies were widely considered and deployed, exactly when they were enacted varied by region. In some places they were unpopular and compliance was inconsistent. In other places, private groups enacted WFH policies even without a government mandate. As a result, the actual times changes in behavior happened are unclear.
We have developed Covid-WFH Detection Algorithms that infer WFH through measuring Internet use. We hope these algorithms can help understand how our world has reacted (and continues to react) to Covid-19.
Our Covid-WFH algorithms are described in a technical report. To summarize:
We begin with data from “pinging” the world. We use data from Trinocular, our Internet outage detection system. As of 2020 it is scanning about 5M /24 IPv4 adddress blocks every 11 minutes. We gather data from multiple observes and post-process this data to estimate the current state of each address.
We identify change-sensitive blocks in this data. A block is change-sensitive if we observe regular daily changes in it (that is, it is diurnal, and if those changes are large enough that we believe we can tell when they change. We see about 250k change-sensitive blocks.
We de-trend this data. The currently active addresses are a combination of daily and weekly patterns over an underlying trend. We fit a mathematical model ot the data so we an extract this trend from the “noise” of regular patterns.
We look for downward trends in change-sensitive blocks. A downward trend indicates a drop in the number of daily users of that block, after we filter out the daily and weekly trends. We believe that downward trends indicate the start of work-from-home, because the the work-day increase and decrease in active addresses of people goes away.
Our approach is described in our technical report: 
We identified several specific downtrend events that correspond to Covid-WFH changes, as well as some downtrend events that are actual changes in Internet usages but are caused by non-Covid-related lockdowns. We describe tehse in our technical report.