add cutoff parameter for train

MoritzM00 · Nov 14, 2024 · d2526d4 · d2526d4
1 parent 08add1e
commit d2526d4
Show file tree

Hide file tree

Showing 3 changed files with 49 additions and 43 deletions.
diff --git a/dvc.lock b/dvc.lock
@@ -102,8 +102,8 @@ stages:
       nfiles: 7
     - path: src/probafcst/pipeline/train.py
       hash: md5
-      md5: 8c38b7144ddc190301187bcbccab381c
-      size: 1003
+      md5: 7d5682f17d005a1324ebe0272cd2bb08
+      size: 1039
     params:
       params.yaml:
         data.energy:
@@ -117,7 +117,8 @@ stages:
         - 0.75
         - 0.975
         train.energy:
-          selected: xgboost
+          selected: quantreg
+          cutoff: '2021-11-14'
           benchmark:
             n_weeks: 75
           quantreg:
@@ -131,8 +132,8 @@ stages:
     outs:
     - path: models/energy_model.pkl
       hash: md5
-      md5: 45a947eb418dca8f2508f4448d4a843d
-      size: 13806992
+      md5: 096054215456ac5e6c7e7858efbda652
+      size: 2764139
   train@bikes:
     cmd: python src/probafcst/pipeline/train.py --target bikes
     deps:
@@ -147,8 +148,8 @@ stages:
       nfiles: 7
     - path: src/probafcst/pipeline/train.py
       hash: md5
-      md5: 8c38b7144ddc190301187bcbccab381c
-      size: 1003
+      md5: 5b991e4c34d3f7c3c0e4e3153ed4d50c
+      size: 1091
     params:
       params.yaml:
         data.bikes:
@@ -162,7 +163,8 @@ stages:
         - 0.75
         - 0.975
         train.bikes:
-          selected: xgboost
+          selected: quantreg
+          cutoff: '2021-11-14'
           benchmark:
             n_weeks: 125
           quantreg:
@@ -176,8 +178,8 @@ stages:
     outs:
     - path: models/bikes_model.pkl
       hash: md5
-      md5: 5a18f0021519963b0e0a5666e915e982
-      size: 2007180
+      md5: f072e90794c21a2066cd284dea8e6ff8
+      size: 113630
   train@no2:
     cmd: uv run python src/probafcst/pipeline/train.py --target no2
     deps:
@@ -209,12 +211,12 @@ stages:
     deps:
     - path: models/bikes_model.pkl
       hash: md5
-      md5: 5a18f0021519963b0e0a5666e915e982
-      size: 2007180
+      md5: f072e90794c21a2066cd284dea8e6ff8
+      size: 113630
     - path: models/energy_model.pkl
       hash: md5
-      md5: 45a947eb418dca8f2508f4448d4a843d
-      size: 13806992
+      md5: 096054215456ac5e6c7e7858efbda652
+      size: 2764139
     - path: src/probafcst//plotting.py
       hash: md5
       md5: 482a42cf8b0b9196d98b0d8e772d83d2
@@ -242,16 +244,16 @@ stages:
     outs:
     - path: output/bikes_forecast.png
       hash: md5
-      md5: e3c07dd456fa87dc28835445d5b42140
-      size: 63372
+      md5: f9734857202bee86f68c717c96eb60ad
+      size: 55660
     - path: output/energy_forecast.png
       hash: md5
-      md5: 0030d0ed824c242ea65fe72b70a452d0
-      size: 70485
+      md5: a1cf54f1601b2d029c3bc43f48906615
+      size: 69911
     - path: output/submission.csv
       hash: md5
-      md5: 171f4c2aa910e3ef29a5348aa0fc0998
-      size: 1609
+      md5: a45024b08f350c39c893886b337f8686
+      size: 1564
   eval@energy:
     cmd: python src/probafcst/pipeline/evaluate.py --target energy
     deps:
@@ -261,8 +263,8 @@ stages:
       size: 1325417
     - path: models/energy_model.pkl
       hash: md5
-      md5: 45a947eb418dca8f2508f4448d4a843d
-      size: 13806992
+      md5: 096054215456ac5e6c7e7858efbda652
+      size: 2764139
     - path: src/probafcst//backtest.py
       hash: md5
       md5: 30b4d59467717a9db02aef2dc16a3ff5
@@ -291,20 +293,20 @@ stages:
     outs:
     - path: output/energy_eval_results.csv
       hash: md5
-      md5: 51ccd4ba00e7c3c7f9fcde93e2095037
-      size: 4826
+      md5: 28a62f592b9a3f1f24dcae579de45e8f
+      size: 4768
     - path: output/energy_metrics.json
       hash: md5
-      md5: 7ef50c23f2abc167c17d15ba6cee5002
-      size: 180
+      md5: ae2bf7575652ca03bcc7d51420580213
+      size: 181
     - path: output/energy_pinball_losses.png
       hash: md5
-      md5: 9352193e2e8a015b2bda9545b951346d
-      size: 14698
+      md5: 228cc9e24fde90a4f586a8ed3d4eea9e
+      size: 13423
     - path: output/eval_plots/energy/
       hash: md5
-      md5: f96a28c200f6bd1cdd999bbac202f432.dir
-      size: 430647
+      md5: 6f406e5b7ef1f7794a0812d817c3d2d4.dir
+      size: 369507
       nfiles: 3
   eval@bikes:
     cmd: python src/probafcst/pipeline/evaluate.py --target bikes
@@ -315,8 +317,8 @@ stages:
       size: 63848
     - path: models/bikes_model.pkl
       hash: md5
-      md5: 5a18f0021519963b0e0a5666e915e982
-      size: 2007180
+      md5: f072e90794c21a2066cd284dea8e6ff8
+      size: 113630
     - path: src/probafcst//backtest.py
       hash: md5
       md5: 30b4d59467717a9db02aef2dc16a3ff5
@@ -345,18 +347,18 @@ stages:
     outs:
     - path: output/bikes_eval_results.csv
       hash: md5
-      md5: 09e1b262562c73716c096dfcd0c1fce7
-      size: 19368
+      md5: fee8a6e27b155999c721a50ad07497cb
+      size: 19482
     - path: output/bikes_metrics.json
       hash: md5
-      md5: 5bd84f0c97eeaff71318d72810a4c736
-      size: 182
+      md5: 66958376e69a5f37010da45b502e4fc2
+      size: 183
     - path: output/bikes_pinball_losses.png
       hash: md5
-      md5: 19ced770898caac13014027764ef69e9
-      size: 21920
+      md5: 34067e344f4e2b98105e6ae80ea0d746
+      size: 19683
     - path: output/eval_plots/bikes/
       hash: md5
-      md5: f25902814a56909f3be4f081b57e1264.dir
-      size: 227088
+      md5: 750cdffd89426e605df50ebced92dd24.dir
+      size: 205991
       nfiles: 3
diff --git a/params.yaml b/params.yaml
@@ -29,7 +29,8 @@ data:
 
 train:
   bikes:
-    selected: xgboost
+    selected: quantreg
+    cutoff: '2021-11-14' # use only data onwards this date for training the final model
     benchmark:
       n_weeks: 125
     quantreg:
@@ -43,7 +44,8 @@ train:
 
 
   energy:
-    selected: xgboost
+    selected: quantreg
+    cutoff: '2021-11-14'
     benchmark:
       n_weeks: 75
     quantreg:

diff --git a/src/probafcst/pipeline/train.py b/src/probafcst/pipeline/train.py
@@ -28,7 +28,9 @@ def train(target):
     y = pd.read_parquet(data_path).asfreq(params.data[target].freq)
 
     forecaster = get_model(params=params.train[target], quantiles=params.quantiles)
-    forecaster.fit(y)
+
+    y_subset = y.loc[: params.train[target].cutoff]
+    forecaster.fit(y_subset)
 
     # Save the model
     model_dir = Path(params.model_dir)