ML
Isolation Forest for sensor anomalies: how it works and where it misleads
Read flagged rows in a sensor table by their isolation score, understand what contamination does in the model, and see where the method misleads you.
Baselines in forecasting projects: why the naive method must be beaten
A forecasting model does not earn its complexity unless it measurably beats a method as simple as “repeat the last value”.
In short
- A naive forecast carries the last observation forward, and on many industrial series this simple method is surprisingly strong.
- Random shuffling leaks knowledge of the future into training, so a time-ordered method should train on the past and test on the period just after.
- Compare a proposed model with the baseline over the same time windows, and prefer the simple method when the gap is within measurement uncertainty.
Simple methods
A naive forecast carries the last observation forward. A seasonal naive forecast uses the value from the previous season, for example the same hour a week earlier. A moving average is the mean of the last few values. These methods need no setup and are surprisingly strong on many industrial series.
Validation that respects time
Splitting a time series by random shuffling leaks knowledge of the future into training. Evaluate with a method that moves forward in time: train on the past, test on the period right after, advance the window and repeat. With this method, every forecast is made only with data known at that moment.
Choose the error measure
Absolute error depends on the series scale. To compare different series, a scale-free measure relative to the baseline’s error can be used. Percentage errors become undefined near zero. Tie the measure to the cost of the decision: over- and under-forecasting can carry different costs.
Decision rule
Compare the proposed model with the baseline over the same time windows. If the difference is not larger than the measurement uncertainty, use the simple method. This is a suggested evaluation approach; a method’s success depends on the specific data and no field-performance claim is made here.
Frequently asked
What simple baselines should a forecasting project beat first?
A naive forecast carries the last observation forward. A seasonal naive forecast uses the value from the previous season, such as the same hour a week earlier. A moving average is the mean of the last few values. These methods need no setup and are often strong on industrial series.
Which error measure should I use for a forecast baseline comparison?
Absolute error depends on the series scale, so a scale-free measure relative to the baseline’s error can compare different series. Percentage errors become undefined near zero. Tie the measure to the cost of the decision, since over-forecasting and under-forecasting can carry different costs.
Sources and further reading
Pilot steps in this guide are proposed engineering workflows. Examples do not constitute measured field performance.
Questions on this topic
Keep reading.
Full library
ML
Multivariate process monitoring with Hotelling T² and SPE
Read correlated sensors through T² and SPE together rather than one statistic, interpret contribution plots, and learn where the method’s assumptions break.
ML
Without fault labels: unsupervised machine learning in industry
Learn what unlabelled sensor history can and cannot tell you, how to evaluate a detector without labels, and why operating modes need attention first.