AI for Time Series and Anomaly Detection
Table 2.
Summary of Empirical Results
|
| Model Category |
Representative Models |
Primary Strengths |
Weaknesses / Limitations |
Average Performance (F1 / RMSE) |
|
|
| Traditional Statistical |
ARIMA, Holt-Winters |
Interpretable, low complexity |
Poor scalability, weak with nonlinear data |
F1 ≈ 0.60 / RMSE ↑ 15–20% |
|
| Machine Learning |
SVM, Random Forest, XGBoost |
Moderate accuracy, interpretable |
Heavy feature engineering |
F1 ≈ 0.75 / RMSE ↓ 10% |
|
| Deep Sequential |
LSTM, GRU |
Captures temporal dependencies |
Slow training, gradient issues |
F1 ≈ 0.85 / RMSE ↓ 18% |
|
| Deep Convolutional |
TCN |
Fast inference, robust to noise |
Limited long-term context |
F1 ≈ 0.88 / RMSE ↓ 20% |
|
| Transformer-Based |
TFT, Informer, TimesNet |
High accuracy, interpretable via attention |
Computationally expensive |
F1 ≈ 0.91 / RMSE ↓ 25% |
|
| Generative / Hybrid |
Autoencoder, VAE, GAN |
Excellent anomaly detection |
Hard to tune, interpretability issues |
F1 ≈ 0.93 / RMSE ↓ 22% |
|
|
|
