I'm training a model to predict percentage change in prices. Both MSE and RMSE are giving me up to 99% accuracy but when I check how often both actual and prediction are pointing in the same direction ((actual >0 and pred > 0) or (actual < 0 and pred < 0)), I get about 49%.

Please how do I define a custom loss that penalizes opposite directions very heavily. I'd also like to add a slight penalty for when the predictions exceeds the actual in a given direction.

So

• actual = 0.1 and pred = -0.05 should be penalized a lot more than actual = 0.1 and pred = 0.05,
• and actual = 0.1 and pred = 0.15 slightly more penalty than actual = 0.1 and pred = 0.05

CodePudding user response：

I will leave it up to you to define your exact logic, but here is how you can implement what you want with tf.cond:

import tensorflow as tf

y_true = [[0.1]]
y_pred = [[0.05]]
mse = tf.keras.losses.MeanSquaredError()

def custom_loss(y_true, y_pred):
  penalty = 20

  # actual = 0.1 and pred = -0.05 should be penalized a lot more than actual = 0.1 and pred = 0.05
  loss = tf.cond(tf.logical_and(tf.greater(y_true, 0.0), tf.less(y_pred, 0.0)),
                   lambda: mse(y_true, y_pred) * penalty,
                   lambda: mse(y_true, y_pred) * penalty / 4)
  
  #actual = 0.1 and pred = 0.15 slightly more penalty than actual = 0.1 and pred = 0.05
  loss = tf.cond(tf.greater(y_pred, y_true),
                   lambda: loss * penalty / 2,
                   lambda: loss * penalty / 3)
  return loss 
  
print(custom_loss(y_true, y_pred))