idea of timedelta based algorithm

src/delta_barth/analysis/forecast.py

@@ -8,6 +8,7 @@ from dataclasses import asdict
 from datetime import datetime as Datetime
 from typing import TYPE_CHECKING, Final, TypeAlias, cast
 
+import dopt_basics.datetime
 import numpy as np
 import pandas as pd
 import scipy.stats
@@ -15,6 +16,7 @@ import sqlalchemy as sql
 
 # --- new: for calculating timedelta
 from dateutil.relativedelta import relativedelta
+from dopt_basics.datetime import TimeUnitsTimedelta
 from sklearn.metrics import mean_absolute_error, r2_score
 from sklearn.model_selection import KFold, RandomizedSearchCV
 from xgboost import XGBRegressor
@@ -251,26 +253,57 @@ def _process_sales(
     # TODO: therefore, stepping with fixed value n does not result in timedelta of n episodes
     # Option A: pad data frame with zero values --> could impede forecast algorithm
     # Option B: calculate next index based on timedelta
-    dates = monthly_sum.index
+    stride = dopt_basics.datetime.timedelta_from_val(365, TimeUnitsTimedelta.DAYS)
+
+    dates = cast(pd.DatetimeIndex, monthly_sum.index)
+
+    min_date = dates.min()
+
     # print("dates: ", dates)
     # ?? --- new: use monthly basis for time windows
     # baseline: 3 years - 36 months
     starting_date = datetime.datetime.now() - relativedelta(months=12)
     # starting_date = dates.max() - relativedelta(months=36)
-    start_index = next(
+    # start_index = next(
-        (i for i, date in enumerate(dates) if date >= starting_date), len(dates) - 1
+    #     (i for i, date in enumerate(dates) if date >= starting_date), len(dates) - 1
-    )
+    # )
-    print("start idx: ", start_index, "length dates: ", len(dates))
+    # print("start idx: ", start_index, "length dates: ", len(dates))
 
-    for add_year, date_idx in enumerate(range(start_index, -1, -12)):
+    def get_index_date(
-        print("date_idx: ", date_idx)
+        dates: pd.DatetimeIndex,
-        first_date = dates[date_idx]
+        starting_date: datetime.datetime | pd.Timestamp,
-        print("first date: ", first_date)
+    ) -> tuple[pd.Timestamp, bool]:
+        target, succ = next(
+            ((date, True) for date in dates if date >= starting_date), (dates[-1], False)
+        )
+        return target, succ
+
+    first_date, succ = get_index_date(dates, starting_date)
+    if not succ:
+        # !! return early
+        ...
+
+    date_span = first_date - min_date
+    steps = date_span.days // stride.days
+
+    for step in range(steps + 1):
+        print("step: ", step)
+        target_date = first_date - step * stride
+        print("target date: ", target_date)
         split_date = dates[-6]
 
+        index_date, succ = get_index_date(dates, target_date)
+
+        if not succ:
+            break
+
+        if index_date >= split_date:
+            print("Skip because of date difference")
+            continue
+
         train = cast(
             pd.DataFrame,
-            monthly_sum.loc[first_date:split_date].copy(),  # type: ignore
+            monthly_sum.loc[index_date:split_date].copy(),  # type: ignore
         )
         print(train)
         print("Length train: ", len(train))
@@ -284,7 +317,7 @@ def _process_sales(
         # ?? --- new: adapted condition to fit new for-loop
         # test set size fixed at 6 --> first iteration: baseline - 6 entries
         # for each new year 10 new data points needed
-        if len(train) >= 30 + 10 * add_year:
+        if len(train) >= 30 + 10 * step:
             too_few_month_points = False
 
             rand = RandomizedSearchCV(
@@ -308,7 +341,7 @@ def _process_sales(
                     best_score_mae = error
                     best_score_r2 = cast(float, r2_score(y_test, y_pred))
                     # --- new: use first_date for best_start_year
-                    best_start_year = first_date.year
+                    best_start_year = target_date.year
                     # --- new: store best_estimator
                     best_estimator = copy.copy(rand.best_estimator_)