I joined the Forecast team at the start of August this year as an Analyst after completing a Geosciences PhD. Although my doctoral research was associated with earthquakes, I was mainly handling and analysing earthquake data through computer programming. I had decided on this field of research after my undergraduate studies in physics as I was interested in how data could be harnessed in the forecasting of earthquakes. I had always been fascinated with different types of catastrophic failure, particularly in earth sciences, and so it seemed like an exciting step to be able to look at how data is used to forecast them.
Forecasting is a technique used in many different fields to determine an event in the future based on previous trends. This could be applied to events like the weather, where meteorologists use measurements of temperature, humidity and wind, to determine how the atmosphere will evolve in the future to be able to make weather forecasts. Forecasting can also be used in a business aspect, whereby a company will forecast their sales in order to make decisions on how many products to purchase.
Forecasting the occurrence of catastrophic failures, such as earthquakes, volcanic eruptions and landslides, is a very active topic in the field of geosciences, as what happens prior to these events is still poorly understood. This is where my recent doctoral research came in. I looked at finding similarly shaped small events, measured through the continuously recorded ground motion at seismic stations. These were important as they have previously been observed to occur prior to these failures. Additionally, the similarity in the shape of the events indicate that they come from the same location. Hence, detecting and analysing these small events has the potential to reveal patterns and relationships with the larger failure events. Understanding the small event behaviour can therefore improve the probabilistic forecasts of these failures.
In my research, I developed a pattern matching technique to extract the similar events to already known events, determine if there were any patterns, and then examine their relationship to physical processes. I applied my method to two large earthquake sequences, a swarm sequence (where there are many small earthquakes without a large event), during an active volcanic period, and prior to a large landslide. An example of one of my case studies of known seismicity during 15thMarch – 2ndApril 2014 in the United States Geological Survey’s catalogue prior to a large Magnitude 8.2 earthquake in Iquique, Chile is shown by the circles below. The brown triangle shows the location of the nearby seismic station used in my analysis.
Illustration of the locations and magnitudes of the 166 events in the United States Geological Survey’s catalogue for the Magnitude 8.2 earthquake sequence in Iquique, Chile from 15th March – 2nd April 2014.The mainshock occurred at 23:46 UTC on 1st April 2014. The location of a nearby seismic station used in my analysis is shown by the brown triangle.
The results of my research showed that my method detected many new small events, which were then attributed to certain drivers accompanying the failure in each case studied. An example below shows how the number of events accelerated prior to a landslide in Nuugaatsiaq, Greenland, following an inverse power law. Each event increased the stress on the slope, which in turn drove the increasing event rate to catastrophic failure. In the long term, the new data detected for these case studies will also help quantify their forecasting.
Example of the acceleration in the rate of events prior to a landslide, which followed an inverse power law. Each event increased the stress on the slope, which in turn drove this increasing event rate to catastrophic failure.
The current probabilistic forecasting of earthquakes produces a likelihood of an event of a given size occurring for a region of interest within a given time period. This type of forecasting is useful for building design implementation, determination of insurance policies, and emergency preparedness. It can also be portrayed in a seismic hazard map, which is a forecast of how the seismicity would affect a region.
A parallel can be drawn between the probabilistic forecasting of earthquakes, and the work that the team do at Forecast. Instead of forecasting the likelihood of an earthquake (or other type of failure event), the likelihood of a customer churning can be calculated. A customer churning is a key metric for any B2C business and just as my work was looking for small events related to the endmember of a failure event, variables associated with a customer’s behaviour can be modelled to forecast their churn rate. This could be indicators such as, how many calls has a customer made to a company’s complaints line, or how their previous behaviour changed depending on last year’s price increase.
Whatever the purpose, the timeliness of the forecast is paramount in taking the necessary steps. This could be for authorities to mitigate hazards after an earthquake, or for a company to take out reactive retention tools to customers more likely to churn. Although I am only 6 weeks into my time at Forecast, I’m very excited to be applying my previous work to modelling different scenarios and learning more from the team.