The COVID-19 pandemic has elevated the visibility and importance of data to the global public as health and government institutions interpret and communicate accurate and timely insights about the pandemic using data analytics. The availability of public data on international responses to the COVID-19 pandemic, limited as it may be, has allowed for a more insightful evaluation of lockdown patterns across the world. To gain insights on possible return-to-normalcy scenarios, data scientists at Cyient modeled data to predict a return-to-normalcy timeline and worked on drawing data from the containment policies dataset published by the University of Oxford. Cyient’s efforts are also in support of an initiative through the Emergent Alliance to understand what global recovery looks like after the COVID-19 pandemic.
The Data-Driven Approach
Eight economic indicators have been tracked, including schools, workplaces, public events, gatherings, public transport, work-from-home impositions, internal movement of people, and international travel. Across these eight indices, ranked between 0 and 4, with 0 indicating normal activity and 4 indicating absolute restriction, we have analyzed lockdown decisions implemented by various countries. When interpreted on a timeline, these eight indices provide a level of economic activity (or lack of activity) for the availability of services or the functioning of essential services in a country.
Our initial exploratory data analysis (EDA) indicated that every country had witnessed the spread pattern locally, both in time and impact. The governmental response through economic lockdown actions carried a unique country-specific signature. Data indicates that the rate of growth of COVID-19 cases seems to be the primary factor that influenced lockdown decisions. This finding is important as it supports the hypothesis that lockdown decisions, and their subsequent removal, would be specific to geographical regions with no discernible global pattern observed, either in time or in removal order.
Cyient used the lockdown indices as input streams to our forecasting algorithm based on a neural net derivative technique, referred to as Long Short-Term Memory (LSTM). Since the lockdown data is transient, i.e., new data is available daily with a potential change of rank in one or more economic activity indicators, we have a strong training set to draw from that forms the essential input time series. A time series of sequences consisting of the daily eight-dimensional vector was used as the input feature. Multiple sequences spanning several months of data constitute the training space.
A deep learning method, such as LSTM, is preferred due to its flexibility in analyzing both structured and unstructured data while also providing control of the learning parameters needed for computational efficiency. Leveraging LSTM models available in the Keras package of the Python framework, Cyient predicted and interpreted possible future sequences that might take shape daily. The model uses the current day’s sequence and all sequences from preceding days to predict the sequence for the following day in a feed-forward scheme. Our test accuracy has been approximately 70%, indicating high reliability of the forecasting model.
Insights from Predictive Analytics
Interpreting the results, we hypothesize the beginning of a return to normalcy if at least 50% of the numbers in the eight-dimensional input stream predict a value of 0. In our evaluation, we relied on data provided for New Zealand as the test subject. The model predicted a return to normalcy beginning August 23, 2020, for this geography. Specifically, the model predicts that these indicators,
Holding public events
Restrictions on internal movement
which earlier ranked as Category 2, would be eased to the reflect pre-pandemic levels of Category 0.
Articles such as the August 30, 2020, Time.com piece New Zealand’s Largest City Exits Lockdown After Bringing Mystery COVID-19 Surge Under Control, have provided evidence to support our claim. According to published reports, August 31, 2020, is the date when Level 3 restrictions were removed in New Zealand.
Enhancing this initial finding, Cyient has formulated another LSTM model that builds on the output of the first model. Using the predicted sequence of the first model as the pivot or primary sequence, we have synthesized several permutations ensuring that the overall count of Category 0 rankings, i.e., normal indices, is identical to the primary sequence across each sequence and in the count. Using this new data set of sequences as input streams to the second LSTM model, Cyient predicts that a complete return to normalcy in New Zealand would likely happen around December 18, 2020. On or around this date, the model predicts complete removal of restrictions across the eight economic activity indicators with workplace restrictions being the last to be eased. We are currently validating the model using a blind test on data from France and Columbia, to test its rigor. We are also identifying the operating boundaries of the model to prescribe its correct usage and interpretation.
As the COVID-19 pandemic has shown, data is becoming more critical to helping experts understand and make recommendations, and for government, businesses, and the public to assess, take action, and plan for the future with information, knowledge, and wisdom founded in data science.
Cyient is a member of the Emergent Alliance, a diverse collaboration of corporates, individuals, NGOs, data specialists, and Governments that contribute expertise, data, and resources to make informed decisions on regional and global economic challenges that will aid societal recovery and help people and businesses thrive post COVID-19.