As said in the previous answer, it depends on data, but also on the algorithm. For example, if you are analysing the time series data or have a long-term dependency, then techniques like CNN do not perform well due to issues such as vanishing gradient. In that case, it also depends on what algorithm you are using (algorithms like LSTM and GRU are designed for that purpose). Further, it also depends on how well you have trained your model.

So if a model is not performing well, then I would suggest (1) check data (how good data is or perform exploratory data analysis to understand data), (2) is your model overfitting or underfitting? Then check your hyperparameter, try to tune it, and (3) check the algorithm - have you selected the correct algorithm?

It's very simple. All forecasting methods are based on the search for patterns in the retrospective data and on the assumption (hypothesis) that these patterns will be valid in the future for the forecast period. In other words, it is assumed that the training sample is representative for a certain period in the future. This period is called the period of ergodicity. But this is an incorrect assumption. Sometimes the patterns in the modeled domain change. The period of ergodicity is violated. New patterns are being formed, although the old ones may remain. Therefore, the point of violation of ergodicity is called the bifurcation point. It is necessary to predict not only based on the patterns of the past period, but also to predict the risks of violating these patterns. I did it back in 1994: http://lc.kubagro.ru/aidos/aidos02/7.4.htm (see Figure 7.2).

https://www.researchgate.net/publication/351924470

Это очень просто. Все методы прогнозирования основаны на поиске закономерностей в ретроспективных данных и на предположении (гипотезе), что эти закономерности будут действовать и в будущем на период прогнозирования. Иначе говоря, предполагается, что обучающая выборка репрезентативна на определенный период в будущее. Этот период называется периодом эргодичности. Но это неверное предположение. Иногда закономерности в моделируемой предметной области меняются. Период эргодичности нарушается. Формируются новые закономерности, хотя могут оставаться и прежние. Поэтому точка нарушения эргодичности называется точка бифуркации. Надо прогнозировать не только основываясь на закономерностях прошлого периода, но и прогнозировать риски нарушения этих закономерностей. Я это делал еще в 1994 году: http://lc.kubagro.ru/aidos/aidos02/7.4.htm (см. рис. 7.2).

https://www.researchgate.net/publication/351924470