If the main goal is to detect whether only the most recent point is anomalous in a univariate time series, your current two-step CNN approach (global + local) is a bit overcomplicated and delicate due to the daily masking/cleaning loop.

Alternatively:

A. Use a rolling forecast error approach:

Develop a model (e.g., ARIMA, LSTM, or even a simple moving average) to predict the next point. Then:

python

Copy-edit

error = abs(actual[-1] - predicted[-1])

is anomaly = error > threshold

Cultivate the limit using a rolling error distribution or confidence interval (e.g., mean + 3*std of past residuals).

B. Statistical test or z-score on residuals:

Establish a baseline model (even just a rolling mean), then for the latest value:

python

CopyEdit

residual = actual[-1] - rolling_mean[-1]

z_score = residual / rolling_std[-1]

is_anomaly = abs(z_score) > 3