# Tag Info

3

One major problem with the MAD/Mean especially in an intermittent demand forecasting context is the following: the MAD will be minimized in expectation by the median of the future distribution. For intermittent data, this may easily be zero. So the "best" forecast, in terms of the MAD/Mean, may be a flat zero line. This is usually not what you want,...

3

In short: Yes, you should tune hyperparameters whenever you add features to a model. By adding features you add extra dimensions for the algorithm to work with and also some noise. In my experience hyperparameters are quite dependent upon the number of features and what information contained in those features. Thus, re-tuning is advised. I don't know of any ...

3

Correct. You have draws from posterior, so for a new $x_i$ the corresponding prediction for $\mu$ is $\dfrac{1}{N} \sum _i \alpha_i + \beta_i x%$. If you're familliar with rstanarm, this is what posterior_linpred does. Since the predictor is linear, this should be equal to $E(\alpha) + E(\beta)x$, where $E$ is the expectation. Here is an example using Stan. ...

2

The width of the prediction interval is probably not the best way of acting as a "warning sign". If you have tons of data, the estimates will be very precise and hence the prediction interval will be relatively constant even when you extrapolate far beyond the original data. See the example below. The training data has 95% of its observations ...

1

You can create a decision tree with dependent features, and they can work very well in the sense that the resulting model is good at predicting new samples. Random forests and XGBoost have been widely used for many years in industry, and competitions too. I would argue this is in part because of their robustness to "data in practice", where the ...

1

'd' in SARIMAX is coming from integrated. For example if a process is integrated of order 1 it means it becomes stationary if you take the first difference. I believe when you define a SARIMAX model in R and if you take 'd' bigger than 0 it will take the difference 'd' times before fitting the model. If you wanna make the data stationary by yourself, such as ...

1

First off, using the prediction in the denominator is a reasonably common idea, see the survey by Green & Tashman (2009). Also, some people use the average of the prediction and the actual, which is commonly called a "symmetric" MAPE (see What is “symmetry” in evaluation metrics). That said, "we penalize underforecasting more than ...

1

Regarding Neural Networks, this topic has been very nicely covered in a recent NeurIPS 2020 paper entitled The Pitfalls of Simplicity Bias in Neural Networks (Shah et al.). I totally recommend reading the paper, I think it is very nicely structured and rigorous. Here is an attempt to summarize its main ideas: The core of the problem is that the field seems ...

Only top voted, non community-wiki answers of a minimum length are eligible