The following content is in the read distance data bosses, technical part of the article summarizes the original please check: rambling time series forecasting – byte articles – zhihu zhuanlan.zhihu.com/p/486343380

1. Effect test

  • Numerical dimension insensitive metric:

    • SMPAE (Symmetric MAPE) :


      • S M A P E = 1 n ∑ t = 1 n ∣ F t – A t ∣ ( A t + F t ) / 2 SMAPE=\frac{1}{n}\sum_{t=1}^{n}\frac{|F_t-A_t|}{(A_t+F_t)/2}

      • The percentage represents the form MAE

      • Upper (200%) lower bound (0%)

    • WMAPE (weighted MAPE) :


      • W M A P E = ∑ t = 1 n ∣ A t – F t ∣ ∑ t – 1 n ∣ A t ∣ WMAPE=\frac{\sum_{t=1}^{n}|A_t-F_t|}{\sum_{t-1}^n|A_t|}

      • Evaluate the errors of different orders of magnitude equivalently

      • Solve MAPE divisor to zero

2. Model classification

2.1 Traditional Methods

  • MA (moving average, averaging the observed values of the past N time points as the prediction of the next point) :

    • Advantages: fast, as baseline effect is good.
    • Disadvantages: can not do many steps, and lag problems.

  • ARIMA (autoregressive moving average) :

    • Advantages: better fit than MA and AR
    • Disadvantages: parameter finding, long running time, need to predict each sequence separately.

  • Prophet (Additive Temporal decomposition model) :

    • Advantages: Non-linear trends have an advantage over seasonal external variables and can output probabilities.
    • Disadvantages: Single sequence prediction, large scale effort.

  • Others: Orbit and NeuralProphet, not so good.

  • Disadvantages:

    • The timing sequence itself has some property requirements, non-end-to-end optimization;
    • Can only do a single sequence prediction, high performance overhead.
    • Autoregressive, no covariable can be introduced.
    • Multi-step prediction doesn’t work.

2.2 ML method

  • The modeling approach

    • Timing conversion table class problems

  • Characteristics of the engineering

    • New features are constructed according to different feature types, and the pretreatment part is not discussed here.

    • Automatic time series feature engineering tool – Tsfresh (According to other documents, full automation is not smart enough to generate too many features and may explode memory)

      • Tsfresh official documentation
      • Tsfresh in Chinese

    • Other features of sequence method is introduced: zhuanlan.zhihu.com/p/67832773

  • model

    • GBDT:

      • LightGBM, Fastai (Optuna or FLAML can be used for auto-tuning)
      • The expression of business characteristics is better than NN

    • NN:

      • Class variable representation learning will have better embedding
      • Flexible loss design
      • Multi-objective learning is more convenient than tree model.

2.3 DL method

  • RNN (RNN, LSTM, GRU)
  • Seq2Seq (RNN combination)

- RNN component as the basic unit, encoder to do training window information extraction, decoder to do prediction of multiple output. - Evaluation: general effect, high calculation cost, poor stability, error analysis and model interpretation difficult to do.Copy the code
  • WaveNet (Hollow causal convolution)

-Better parallelism than RNN, one-dimensional CNN was used for sequence prediction, residual connection and skip connection were added, and a series of complex door mechanisms were added. - Evaluation: Not as good as RNNCopy the code
  • LSTNet

- Evaluation: Inferior to Feature engineering + FAstaiCopy the code
  • DeepAR

- Seq2Seq, but can output the probability distribution. - Evaluation: it is difficult to stabilize convergence, the accuracy fluctuates greatly, and the effect is not better than GBDT.Copy the code
  • N-Beats

- Univariate prediction, with certain explanatory season and trend. - Evaluation: general effect, not good to add additional characteristic variables.Copy the code
  • TFT

- The DL model that can challenge the tree model has certain approximation with the tree model and has characteristic variable selection network. - Evaluation: somewhat interesting, due to the principle of its quasi tree form, the effect is stable, but the calculation cost is still relatively large.Copy the code

  • Conclusion:

    • At present, GBDT is better than DL in large scale
    • DL’s strengths include pre-train, transfer learning and represent learning, and generation model is difficult to be applied to timing domain.

2.4 the sequential AutoML

  • Library:

    • Auto_TS

  • Mainstream methods:

    • Feature engineering +GBDT
    • TFT structure makes it feasible to automate feature engineering of data set interface.

\