Robust and Scalable Bayesian Online Changepoint Detection
Authors: Matias Altamirano, Francois-Xavier Briol, Jeremias Knoblauch
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We investigate Dm-BOCD empirically in several numerical experiments. In doing so, we highlight its computational and inferential advantages over standard BOCD and β-BOCD. In all experiments, we choose conjugate priors as in Proposition 3.1, and m and ω as in Section 3.4. All code and data is publicly available at https://github. com/maltamiranomontero/DSM-bocd. |
| Researcher Affiliation | Academia | 1Department of Statistical Science, University College London, London, United Kingdom 2The Alan Turing Institute, London, United Kingdom. |
| Pseudocode | No | The paper describes the methodology using prose and mathematical equations but does not include any pseudocode or algorithm blocks. |
| Open Source Code | Yes | All code and data is publicly available at https://github. com/maltamiranomontero/DSM-bocd. |
| Open Datasets | Yes | The data is publicly available on First Rate Data.1 That day, the Associated Press Twitter account was hacked and falsely tweeted that explosions at the White House had injured then-president Barack Obama. ... data which is publicly available on Yahoo finance.2 FTT was the token issued by FTX... The data is publicly available via the Bank of England database.3 Since the 10-year yield has been positive throughout history, we model it using the gamma distribution. |
| Dataset Splits | Yes | To operationalise this, we choose ω = arg minω>0 KL πDm ω (θ|x1:t ) πB(θ|x1:t ) . Computing ω is implemented using automatic differentiation via jax (Bradbury et al., 2018). ... In the Twitter flash crash experiment... We use the first 50 observations to select ω as in Section 3.4, with an obtained value of ω 0.0001 ... In the Cryptocrash experiment... we manually fix ω = 0.01. ... For Dm-BOCD, we use a conjugate squared exponential prior r with parameters µ = (0, 1) , and Σ a diagonal matrix so that diag(Σ) = (10, 1) for the natural parameters of said Gaussian. For standard BOCD, we use a Normal-inverse gamma prior with parameters µ0 = 0, ν = 1, α = 2 and β = 10. ... We use the first 200 observations to select ω as in Section 3.4, with an obtained value of ω 0.0004 ... We use the first 100 observations to select ω as in 3.4. The obtained value is ω 0.05. |
| Hardware Specification | Yes | On a machine with processor Intel i7-7500U 2.7 GHz, and 12GB of RAM, Dm-BOCD took about 10 times less than β-BOCD. |
| Software Dependencies | No | Computing ω is implemented using automatic differentiation via jax (Bradbury et al., 2018). (No specific version number for JAX or other libraries is provided). |
| Experiment Setup | Yes | In all experiments, we choose conjugate priors as in Proposition 3.1, and m and ω as in Section 3.4. ... For the Dm-BOCD, we use a conjugate squared exponential prior r with parameters µ = (0, 1) , and Σ a diagonal matrix so that diag(Σ) = (10, 1) for the natural parameters of said Gaussian. For standard BOCD, we use a Normal-inverse gamma prior with parameters µ0 = 0, ν = 1, α = 2 and β = 10. ... For all experiments, we take k = 50. |