fix: X_df handling in direct approach (#468)

Author: jmoralez

mlforecast/core.py

@@ -703,6 +703,7 @@ def _predict_multi(
         result = df_constructor({self.id_col: uids, self.time_col: dates})
         for name, model in models.items():
             with self._backup():
+                self._predict_setup()
                 new_x = self._get_features_for_next_step(X_df)
                 if before_predict_callback is not None:
                     new_x = before_predict_callback(new_x)
@@ -789,10 +790,16 @@ def predict(
                     raise ValueError(
                         f"The following features were provided through `X_df` but were considered as static during fit: {common}.\n"
                         "Please re-run the fit step using the `static_features` argument to indicate which features are static. "
-                        "If all your features are dynamic please pass an empty list (static_features=[])."
+                        "If all your features are dynamic please provide an empty list (static_features=[])."
                     )
                 starts = ufp.offset_times(self.last_dates, self.freq, 1)
-                ends = ufp.offset_times(self.last_dates, self.freq, horizon)
+                if getattr(self, "max_horizon", None) is None:
+                    ends = ufp.offset_times(self.last_dates, self.freq, horizon)
+                    expected_rows_X = len(self.uids) * horizon
+                else:
+                    # direct approach uses only the immediate next timestamp
+                    ends = starts
+                    expected_rows_X = len(self.uids)
                 dates_validation = type(X_df)(
                     {
                         self.id_col: self.uids,
@@ -803,7 +810,7 @@ def predict(
                 X_df = ufp.join(X_df, dates_validation, on=self.id_col)
                 mask = ufp.between(X_df[self.time_col], X_df["_start"], X_df["_end"])
                 X_df = ufp.filter_with_mask(X_df, mask)
-                if X_df.shape[0] != len(self.uids) * horizon:
+                if X_df.shape[0] != expected_rows_X:
                     msg = (
                         "Found missing inputs in X_df. "
                         "It should have one row per id and time for the complete forecasting horizon.\n"

nbs/core.ipynb

@@ -1184,6 +1184,7 @@
     "        result = df_constructor({self.id_col: uids, self.time_col: dates})\n",
     "        for name, model in models.items():\n",
     "            with self._backup():\n",
+    "                self._predict_setup()\n",
     "                new_x = self._get_features_for_next_step(X_df)\n",
     "                if before_predict_callback is not None:\n",
     "                    new_x = before_predict_callback(new_x)\n",
@@ -1261,10 +1262,16 @@
     "                    raise ValueError(\n",
     "                        f\"The following features were provided through `X_df` but were considered as static during fit: {common}.\\n\"\n",
     "                        \"Please re-run the fit step using the `static_features` argument to indicate which features are static. \"\n",
-    "                        \"If all your features are dynamic please pass an empty list (static_features=[]).\"\n",
+    "                        \"If all your features are dynamic please provide an empty list (static_features=[]).\"\n",
     "                    )\n",
     "                starts = ufp.offset_times(self.last_dates, self.freq, 1)\n",
-    "                ends = ufp.offset_times(self.last_dates, self.freq, horizon)\n",
+    "                if getattr(self, 'max_horizon', None) is None:\n",
+    "                    ends = ufp.offset_times(self.last_dates, self.freq, horizon)\n",
+    "                    expected_rows_X = len(self.uids) * horizon\n",
+    "                else:\n",
+    "                    # direct approach uses only the immediate next timestamp\n",
+    "                    ends = starts\n",
+    "                    expected_rows_X = len(self.uids)\n",
     "                dates_validation = type(X_df)(\n",
     "                    {\n",
     "                        self.id_col: self.uids,\n",
@@ -1275,7 +1282,7 @@
     "                X_df = ufp.join(X_df, dates_validation, on=self.id_col)\n",
     "                mask = ufp.between(X_df[self.time_col], X_df['_start'], X_df['_end'])\n",
     "                X_df = ufp.filter_with_mask(X_df, mask)\n",
-    "                if X_df.shape[0] != len(self.uids) * horizon:\n",
+    "                if X_df.shape[0] != expected_rows_X:\n",
     "                    msg = (\n",
     "                        \"Found missing inputs in X_df. \"\n",
     "                        \"It should have one row per id and time for the complete forecasting horizon.\\n\"\n",
@@ -2015,7 +2022,7 @@
       "text/markdown": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L757){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L758){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "## TimeSeries.predict\n",
        "\n",
@@ -2029,7 +2036,7 @@
       "text/plain": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L757){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L758){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "## TimeSeries.predict\n",
        "\n",
@@ -2167,7 +2174,7 @@
       "text/markdown": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L862){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L869){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "## TimeSeries.update\n",
        "\n",
@@ -2180,7 +2187,7 @@
       "text/plain": [
        "---\n",
        "\n",
-       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L862){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
+       "[source](https://github.com/Nixtla/mlforecast/blob/main/mlforecast/core.py#L869){target=\"_blank\" style=\"float:right; font-size:smaller\"}\n",
        "\n",
        "## TimeSeries.update\n",
        "\n",

nbs/forecast.ipynb

@@ -1149,6 +1149,7 @@
     "import numpy as np\n",
     "import xgboost as xgb\n",
     "from sklearn.linear_model import LinearRegression\n",
+    "from utilsforecast.feature_engineering import time_features\n",
     "from utilsforecast.plotting import plot_series\n",
     "\n",
     "from mlforecast.lag_transforms import ExpandingMean, ExponentiallyWeightedMean, RollingMean\n",
@@ -5439,6 +5440,33 @@
     "preds2 = fcst2.predict(10)\n",
     "pd.testing.assert_frame_equal(preds, preds2)"
    ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4a39447b-3d6b-4960-83c9-4f927c472ebb",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "#| hide\n",
+    "# direct approach requires only one timestamp and produces same results for two models\n",
+    "series = generate_daily_series(5)\n",
+    "h = 5\n",
+    "freq = 'D'\n",
+    "train, future = time_features(series, freq=freq, features=['day'], h=h)\n",
+    "models = [LinearRegression(), lgb.LGBMRegressor(n_estimators=5)]\n",
+    "\n",
+    "fcst1 = MLForecast(models=models, freq=freq, date_features=['dayofweek'])\n",
+    "fcst1.fit(train, max_horizon=h, static_features=[])\n",
+    "preds1 = fcst1.predict(h=h, X_df=future)  # extra timestamps\n",
+    "\n",
+    "fcst2 = MLForecast(models=models[::-1], freq=freq, date_features=['dayofweek'])\n",
+    "fcst2.fit(train, max_horizon=h, static_features=[])\n",
+    "X_df_one = future.groupby('unique_id', observed=True).head(1)\n",
+    "preds2 = fcst2.predict(h=h, X_df=X_df_one)  # only needed timestamp\n",
+    "\n",
+    "pd.testing.assert_frame_equal(preds1, preds2[preds1.columns])"
+   ]
   }
  ],
  "metadata": {