Chapter Summaries:
0:00 – Module 8 Intro: Time Series in ThingWorx Analytics
An overview of Module 8, introducing time series data, how it differs from cross-sectional data, and the time series functionality available inside ThingWorx Analytics.
0:22 – Cross-Sectional Data: Order-Independent Features & Examples
Defines cross-sectional data as a snapshot from many subjects at a single point in time, with many features and order-independent rows that are ready for analytics immediately — illustrated with heart disease risk and used-car pricing examples.
1:56 – Time Series Data: Order-Dependent Records & Few Features
Explains that time series data comes from one or a few similar subjects over time with few features, and why record order matters, using a pump sensor example and a flip-book analogy to show that shuffled records produce meaningless chaos.
3:13 – Transforming Data + Time Series Examples (Stock, Temperature)
Covers why raw data must first be transformed before a time series model can be built in ThingWorx Analytics, with predictive examples like forecasting next week's stock price from gross revenue and projecting future average global temperature.
3:53 – Temporal Feature & Time Sampling Interval Metadata
Details the first key metadata configuration — the temporal opt type that timestamps every record — and the required time sampling interval (an integer ≥ 1) that must match the data's time steps or training will error out.
4:59 – Entity ID Feature: Separating Multiple Entities
Introduces the entity ID opt type, which tells ThingWorx Analytics which entity each record belongs to (machine 1, 2, 3, and so on) so the engine can cleanly separate the data of one asset from the next.
5:58 – is_static Flag: Preventing Temporal Pivoting
Describes the new is_static flag, which stops the engine from pivoting out temporal features for a raw feature that rarely changes per entity — with examples such as city or manufacturing recipe.
6:36 – Training vs Validation: Out-of-Time Split & Data Leakage
Contrasts the randomized 80/20 split used for cross-sectional data with automatic out-of-time validation for time series, where the last 20% (Builder default) is held out for testing to prevent data leakage between training and validation.
7:42 – Data Prep: Chronological Order, Interpolation & Binning
Best practices for preparing each entity's data — chronological ordering with no gaps, pre-sorting, interpolation for missing values, integer/long time steps as multiples of the time sampling interval, and equal-size binning to pre-process the data.
8:32 – Lookback & Lookahead Explained
Defines the two core terms: lookback (how many recent records are used to make a prediction) and lookahead (how many time steps into the future the prediction targets).
9:09 – Pivoting Operations: Offset Columns & Goal Variable
Shows how ThingWorx Analytics consolidates a time series trend into one record via pivoting — a lookback of four and lookahead of one produces offset 0–3 columns plus a goal variable set by the lookahead value.
9:48 – Statistical Transformations: Derivatives, Volatility & Feature Counts
Explains the extra columns created by splitting history pivots into non-overlapping subsets in powers of two, computing average, first/second/third derivatives, and volatility (standard deviation) — with the feature-count formula yielding 86 features for two raw features at a lookback of eight.
11:24 – Overfitting Risk & Minimizing Raw Features
Warns that the large number of columns generated by statistical transformations can drive overfitting, underscoring why it's important to include as few raw features as possible in the original data set.
