5 ways to fail with time series
summary
Five concrete pitfalls and fixes: mismatched sampling frequency matters creates artifacts when resampling; naive train/test splits leak future information so use incremental time splits; common KPIs hide directionality so choose task‑specific metrics; models deployed without automation fail because systems need to retrain models continuously at scale (examples: thousands of turbine models daily). Also trust sensor error models and signal‑processing libraries, favor simple linear models on filtered data, and convert timestamps to UTC and use domain knowledge.
AI-written summary · about
chapters (29)
- 0:00Intro
- 0:35Why this talk
- 1:32Examples of timeseries
- 4:43Timecritical data
- 5:34General overview of timeseries
- 7:09Timeseries examples
- 7:27Context
- 11:22Finding algorithms
- 12:01Ignoring the history
- 12:54Autoregressive analysis
- 13:32Training models
- 14:24Kpi
- 15:51Dynamics
- 17:34Assumptions
- 19:16Cross dependencies
- 19:44limited knowledge
- 20:25example
- 20:55a rabbit hole
- 21:13an example
- 22:06cardinality
- 22:32python users
- 23:21big data
- 23:50time series models
- 24:01train a model
- 25:59bonus fail
- 26:32things that can go wrong
- 28:35take home messages
- 29:46thank you
- 30:10Quantifying uncertainty
related talks
30mWhen Good Models Go Bad: The damage caused by wayward models and how to prevent it
12mModel Monitoring: Detecting and Analyzing Data Issues
24mMachine learning and financial time series analysis
27m10 Things That Can Go Wrong with ML Projects
30mFrom Black Box to a Known Quantity: How to Build Predictable, Reliable ML-based
23mHow to fail AI implementations
25mWhy Good Models Fail After Deployment
08mConsiderations for producing the best AI and Machine Learning models