Alyona Galyeva & Dr. Sebastian Werner - 5 ways to fail with time series

timeseries data is everywhere and often forgotten, a rabbit hole

• oversized libraries and ignoring context lead to issues *
• don’t assume models are static, model changing assumption *
• and don’t assume models are not static, but model stays the same *
• dealing with large data sets can cause memory issues *
• data normalizations are important *
• when working with sequence data, boundaries play a big role *
• cardinality matters, e.g., sampling every minute affects results *
• resampling data can lead to trail of actual measurements *
• timestamps are important, ignore or don’t understand *
• predictive maintenance is interesting and valuable *
• when working with sensors, one must consider noise and uptime *
• sampling frequency affects model choice *
• Gaussian error distribution can be misleading *
• data processing can easily obstruct analysis *
• don’t assume autoregressive analysis is correct for all timeframes *
• forecasting accuracy is crucial, especially for money *
• timestamps handled between libraries is important *
• don’t ignore cross dependencies between variables *
• working with sequence data can lead to difficulties with noise and boundaries *
• big data assumes all data is valuable, but that’s not always the case *
• assuming absolute timestamps is crucial *
• predicting human behavior can be challenging, involving cultural and personal impacts *
• assuming the forecast is perfect is not realistic *
• don’t ignore sampling rate, physical constraints, and other influences *
• revisit models frequently to maintain performance *
• choose models wisely, taking into account frequency, noise, and sampling rate *
• consider external factors, like volatility, when modeling *
• Thanks notes: