Skip to content

[Before or 0.4v] Constructing more soft label types in training data #17

Description

@ieatyoursushi

y_bt is a good start which uses encoded y_t all the way to y+t+30 data but there are more ways to encode information in soft labels like using a further timeline (like having a list of y_30, y_90, y_180, y* encoded with past and future data ect) the earlier data avaialble, perhaps encoding economic meanings, tax alpha states in a generated DI portfolio, ect, a more infomration-encoded objecive than the current y_bt for 0.1-0.2v. (make sure that feature space does not encode this information on an each lot model treats as iid)

for ex:

$y_{\text{persist},30} = \frac{1}{30} \sum_{s=1}^{30} f^*(x_{t+s})$

$y_{\text{alpha},30} = \frac{1}{30} \sum_{s=1}^{30} \alpha_{\text{tax}}(x_{t+s}) f^*(x_{t+s})$

$y_{\text{max},30} = \max_{1 \leq s \leq 30} \alpha_{\text{tax}}(x_{t+s}) f^*(x_{t+s})$

encodes fairly accurate assumptive behaviors about the stable positions in indicies like the S&P 500 mainly of the fact that high volatility / paradigm shift based price movements are extremely rare while overall trends in a direction over weeks or months is much more common, highest ROI essentailly for the most simplicity.

Each label will have its own leader board of auc/pr scores and such along loss and objective functions towards each label.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions