Timeline for Why does my neural network consistently predict values in the wrong range dispite training having slowed to a stop? [duplicate]
Current License: CC BY-SA 4.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Nov 4, 2022 at 14:24 | comment | added | Stephan Kolassa | It depends on what you mean by "over-/underestimation". The MSE (and scaled variants) elicits the conditional expectation. If you want the conditional median, you can use the MAE: stats.stackexchange.com/q/355538/1352 | |
| Nov 4, 2022 at 13:56 | history | closed |
Stephan Kolassa CommunityBot |
Duplicate of What are the shortcomings of the Mean Absolute Percentage Error (MAPE)? | |
| Nov 4, 2022 at 13:55 | comment | added | Ollie-fork | @StephanKolassa Oh wow that makes so much sense! I'm going to have a look into scaling all of my data better than it currently is done and hopefully that will allow MSE to work properly. Are there any other loss functions that have benefits over MSE that dont prioritise under or overestimation? | |
| Nov 4, 2022 at 13:44 | comment | added | Stephan Kolassa | The MAPE rewards underpredictions, see the proposed duplicate. Of course, the effect is strongest for small actuals, which is consistent with the underpredictions at small values in the top left, top right and bottom left panels. If you want unbiased expectation predictions, use an error measure that is optimized by an unbiased expectation prediction, i.e., the MSE, possibly scaled. | |
| S Nov 4, 2022 at 13:03 | review | First questions | |||
| Nov 4, 2022 at 13:42 | |||||
| S Nov 4, 2022 at 13:03 | history | asked | Ollie-fork | CC BY-SA 4.0 |