Limitations of Machine Learning – What the Forecast App Can't Do

Although the Forecast App is based on powerful LSTM models, it is—like any machine learning method—only as good as the data it works with. This chapter highlights typical pitfalls and misconceptions that can lead to poor or misleading results when using the app.


1. No Forecast Without a Pattern

Problem: If your data contains no recognizable pattern, the model cannot find one either.

Example: A target attribute fluctuates purely randomly or contains no relevant correlation to the input values—in this case, the Forecast App can only estimate average values, but cannot make reliable predictions.

Recognizable by:

  • Flat prediction curves

  • Low model improvement after retraining

  • Very high or very low confidence interval in the output

Solution:

  • First, check in Analytics & Reports whether recognizable trends or patterns exist.

  • Avoid binary, random, or irregular signals without explainable dependencies.


2. Extreme Values and Outliers Disrupt the Model

Problem: A single outlier (e.g., due to a faulty measurement) can severely distort the model structure and prediction behavior.

Example: If all values are in the range of 0.01, but a data point suddenly amounts to 10,000,000, the model will artificially raise future predictions to account for possible "similar" outliers.

Solution:

  • Clean historical data before training (e.g., with filters or manual removal).

  • If necessary, use logarithmic or normalized values.

  • Use the Eliona Calculator to derive differentiated or smoothed values.


3. New Value Ranges Lead to Inaccurate Forecasts

Problem: If values suddenly occur after training that the model has never seen before, it cannot learn a sensible response to them.

Example: A sensor provides values between 10–50 during training. After training, new values in the range of 100–200 arrive. The model does not "know" this range and behaves unpredictably.

Solution:

  • Instead of predicting absolute values, calculate relative changes or differences.

  • In case of changes in data behavior: Retrain the model.


4. Models Need Sufficient Data

Problem: If there is too little data, a viable model cannot be created.

Recommendation:

  • At least several hundred data points should be available.

  • Ideally: Histories with seasonal or periodic fluctuations over several cycles.

The system automatically checks the data length and only starts training if there are enough measured values.


6. Frequent Structural Breaks in the Data History

Problem: Sudden changes (e.g., system changes, changes in measurement methods) lead to structural breaks that the model cannot explain.

Solution:

  • Segment the dataset before training.

  • Avoid mixing different data sources or measurement methods in a forecast.


Conclusion

The Forecast App is a powerful tool—but only with sensibly prepared and structured data. It does not recognize meaning, but only statistical patterns.

The model cannot:

  • Independently recognize faulty data

  • "Understand" or interpret decisions

  • Deal with completely unknown data ranges in a meaningful way

  • Make predictions if no explainable patterns exist

But the model can:

  • Recognize regularities

  • Derive reliable trends from historical patterns

  • Provide forecasts based on consistent, cleaned data

Last updated