- ☀️Daily Log:
- bitcoin was suppose to be a steady form of asset that resists broader economic trends, but so far it has been a speculative tech investment
- in 2022, it has closely mirrored the NASDAQ movements
- #til2022 Sahara Desert Circles
- 🌐:
- 💁♂️: Circle markings in the Sahara Desert in Algeria. Could be a result of seismic surveys used for natural gas extraction
- 🤔:
- #data-science Spatio-Temporal Extreme Event Modelling of Terror Insurgencies
- Authors: Lekha Patel
- Highly unpredictable but high impact outcome events are important to predict accurately
- Arbitrary space-time region
- Inhomogeneous baseline intensity
- Self-exciting marked spatio-temporal model
- Triggering intensity is modelled with Gaussian Process prior distribution to flexibly capture intricate spatio-temporal dependencies between an arbitrary event and previous events
- Novel generalized zipf distribution to measure the intensity of the event
- customized Markov chain Monte Carlo method to estimate the model parameters
- Extreme value analysis focuses on the characterization of events that lies at the tails of a distribution
- i.e. earthquakes, hurricanes, flooding, wildfires
- Two common approaches from statistical point of view
- calculating a sequence of maximum (minimum) values over blocks of data and fitting these values to their large sample distribution (the Generalized extreme value distribution)
- find observations that exceeds (fall below) a given threshold and fits the Generalized pareto distribution to these exceedance values (Peaks over Threshold method (PoT))
- A novel approach is to use point process modelling
- Used to probabilstically describe an event that happens in space-time
- Homogeneous point process can be used if the event occurs in a constant rate
- Self-exciting processes can be used to capture dependencies that naturally arise in events that may have connections
- To measure the different impact of these events, layered a mixture model to estimate the discrete mark of the event
- Marks above a threshold can be modelled with a Generalized-Zipf distribution
- ==It is similar to a two-hurdle model except the first hurdle is the self-exciting point process and the second hurdle is a generalized zipf model==
- Retrospective::