All algorithms have inductive biases which determine what they will be good or bad at. Which model is good on a set of data is determined by the data, not the model. For instance, in classification, Support Vector Machine works wonders on linearly separable data. Whether or not the classes are linearly separable and/or by which features is born out by the data.
It's best to start with a bit of data exploration to see which features are important for prediction and one's favorite "simple" model (i.e. Decision Tree Regression) just to see if decent performance can be simply achieved.
Lastly, just because something has dependency in the time domain does not mean one has to jump to something complex like LSTM. Feel free to include values from previous time steps as features to the present prediction. If it's a predictor for future values K-nearest neighbor is as likely a candidate as the next.