feat: restrict input forecast

Nixtla · Apr 25, 2024 · 603c666 · 603c666
1 parent 1e0b0a8
commit 603c666
Show file tree

Hide file tree

Showing 2 changed files with 26 additions and 9 deletions.
diff --git a/nbs/nixtla_client.ipynb b/nbs/nixtla_client.ipynb
@@ -617,12 +617,23 @@
     "        # - we dont have exogenous regegressors\n",
     "        # - and we dont want to produce pred intervals\n",
     "        # - no add history\n",
-    "        restrict_input = self.finetune_steps == 0 and X_df is None and self.level is not None and not add_history\n",
+    "        restrict_input = self.finetune_steps == 0 and X_df is None and not add_history\n",
     "        if restrict_input:\n",
     "            # add sufficient info to compute\n",
     "            # conformal interval\n",
     "            main_logger.info('Restricting input...')\n",
-    "            new_input_size = 3 * self.input_size + max(self.model_horizon, self.h)\n",
+    "            if self.level is not None:\n",
+    "                # @AzulGarza\n",
+    "                #  this is an old opinionated decision\n",
+    "                #  about reducing the data sent to the api\n",
+    "                #  to reduce latency when\n",
+    "                #  a user passes level. since currently the model\n",
+    "                #  uses conformal prediction, we can change a minimum\n",
+    "                #  amount of data if the series are too large\n",
+    "                new_input_size = 3 * self.input_size + max(self.model_horizon, self.h)\n",
+    "            else:\n",
+    "                # we only want to forecast\n",
+    "                new_input_size = self.input_size\n",
     "            Y_df = Y_df.groupby(\"unique_id\").tail(new_input_size)\n",
     "            if X_df is not None:\n",
     "                X_df = X_df.groupby(\"unique_id\").tail(\n",

diff --git a/nixtla/nixtla_client.py b/nixtla/nixtla_client.py
@@ -559,17 +559,23 @@ def forecast(
         # - we dont have exogenous regegressors
         # - and we dont want to produce pred intervals
         # - no add history
-        restrict_input = (
-            self.finetune_steps == 0
-            and X_df is None
-            and self.level is not None
-            and not add_history
-        )
+        restrict_input = self.finetune_steps == 0 and X_df is None and not add_history
         if restrict_input:
             # add sufficient info to compute
             # conformal interval
             main_logger.info("Restricting input...")
-            new_input_size = 3 * self.input_size + max(self.model_horizon, self.h)
+            if self.level is not None:
+                # @AzulGarza
+                #  this is an old opinionated decision
+                #  about reducing the data sent to the api
+                #  to reduce latency when
+                #  a user passes level. since currently the model
+                #  uses conformal prediction, we can change a minimum
+                #  amount of data if the series are too large
+                new_input_size = 3 * self.input_size + max(self.model_horizon, self.h)
+            else:
+                # we only want to forecast
+                new_input_size = self.input_size
             Y_df = Y_df.groupby("unique_id").tail(new_input_size)
             if X_df is not None:
                 X_df = X_df.groupby("unique_id").tail(