Upload 2 files

Benchmark_QRT_aw81ejz/Benchmark QRT.ipynb

@@ -0,0 +1,688 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Benchmark QRT\n",
+    "\n",
+    "This notebook illustrates a simple benchmark example that should help novice participants to start the competition.\n",
+    "\n",
+    "## Used libraries"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import seaborn as sns\n",
+    "import numpy as np\n",
+    "import pandas as pd\n",
+    "from sklearn.ensemble import RandomForestClassifier\n",
+    "from sklearn.metrics import accuracy_score\n",
+    "from sklearn.model_selection import KFold"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading data\n",
+    "\n",
+    "The train and test inputs are composed of 46 features.\n",
+    "\n",
+    "The target of this challenge is `RET` and corresponds to the fact that the **return is in the top 50% of highest stock returns**.\n",
+    "\n",
+    "Since the median is very close to 0, this information should not change much with the idea to predict the sign of the return."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>DATE</th>\n",
+       "      <th>STOCK</th>\n",
+       "      <th>INDUSTRY</th>\n",
+       "      <th>INDUSTRY_GROUP</th>\n",
+       "      <th>SECTOR</th>\n",
+       "      <th>SUB_INDUSTRY</th>\n",
+       "      <th>RET_1</th>\n",
+       "      <th>VOLUME_1</th>\n",
+       "      <th>RET_2</th>\n",
+       "      <th>VOLUME_2</th>\n",
+       "      <th>...</th>\n",
+       "      <th>VOLUME_16</th>\n",
+       "      <th>RET_17</th>\n",
+       "      <th>VOLUME_17</th>\n",
+       "      <th>RET_18</th>\n",
+       "      <th>VOLUME_18</th>\n",
+       "      <th>RET_19</th>\n",
+       "      <th>VOLUME_19</th>\n",
+       "      <th>RET_20</th>\n",
+       "      <th>VOLUME_20</th>\n",
+       "      <th>RET</th>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>ID</th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>0</th>\n",
+       "      <td>0</td>\n",
+       "      <td>2</td>\n",
+       "      <td>18</td>\n",
+       "      <td>5</td>\n",
+       "      <td>3</td>\n",
+       "      <td>44</td>\n",
+       "      <td>-0.015748</td>\n",
+       "      <td>0.147931</td>\n",
+       "      <td>-0.015504</td>\n",
+       "      <td>0.179183</td>\n",
+       "      <td>...</td>\n",
+       "      <td>0.630899</td>\n",
+       "      <td>0.003254</td>\n",
+       "      <td>-0.379412</td>\n",
+       "      <td>0.008752</td>\n",
+       "      <td>-0.110597</td>\n",
+       "      <td>-0.012959</td>\n",
+       "      <td>0.174521</td>\n",
+       "      <td>-0.002155</td>\n",
+       "      <td>-0.000937</td>\n",
+       "      <td>True</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>1</th>\n",
+       "      <td>0</td>\n",
+       "      <td>3</td>\n",
+       "      <td>43</td>\n",
+       "      <td>15</td>\n",
+       "      <td>6</td>\n",
+       "      <td>104</td>\n",
+       "      <td>0.003984</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>-0.090580</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>...</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>0.003774</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>-0.018518</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>-0.028777</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>-0.034722</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>True</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>2</th>\n",
+       "      <td>0</td>\n",
+       "      <td>4</td>\n",
+       "      <td>57</td>\n",
+       "      <td>20</td>\n",
+       "      <td>8</td>\n",
+       "      <td>142</td>\n",
+       "      <td>0.000440</td>\n",
+       "      <td>-0.096282</td>\n",
+       "      <td>-0.058896</td>\n",
+       "      <td>0.084771</td>\n",
+       "      <td>...</td>\n",
+       "      <td>-0.010336</td>\n",
+       "      <td>-0.017612</td>\n",
+       "      <td>-0.354333</td>\n",
+       "      <td>-0.006562</td>\n",
+       "      <td>-0.519391</td>\n",
+       "      <td>-0.012101</td>\n",
+       "      <td>-0.356157</td>\n",
+       "      <td>-0.006867</td>\n",
+       "      <td>-0.308868</td>\n",
+       "      <td>False</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>3</th>\n",
+       "      <td>0</td>\n",
+       "      <td>8</td>\n",
+       "      <td>1</td>\n",
+       "      <td>1</td>\n",
+       "      <td>1</td>\n",
+       "      <td>2</td>\n",
+       "      <td>0.031298</td>\n",
+       "      <td>-0.429540</td>\n",
+       "      <td>0.007756</td>\n",
+       "      <td>-0.089919</td>\n",
+       "      <td>...</td>\n",
+       "      <td>0.012105</td>\n",
+       "      <td>0.033824</td>\n",
+       "      <td>-0.290178</td>\n",
+       "      <td>-0.001468</td>\n",
+       "      <td>-0.663834</td>\n",
+       "      <td>-0.013520</td>\n",
+       "      <td>-0.562126</td>\n",
+       "      <td>-0.036745</td>\n",
+       "      <td>-0.631458</td>\n",
+       "      <td>False</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>4</th>\n",
+       "      <td>0</td>\n",
+       "      <td>14</td>\n",
+       "      <td>36</td>\n",
+       "      <td>12</td>\n",
+       "      <td>5</td>\n",
+       "      <td>92</td>\n",
+       "      <td>0.027273</td>\n",
+       "      <td>-0.847155</td>\n",
+       "      <td>-0.039302</td>\n",
+       "      <td>-0.943033</td>\n",
+       "      <td>...</td>\n",
+       "      <td>-0.277083</td>\n",
+       "      <td>-0.012659</td>\n",
+       "      <td>0.139086</td>\n",
+       "      <td>0.004237</td>\n",
+       "      <td>-0.017547</td>\n",
+       "      <td>0.004256</td>\n",
+       "      <td>0.579510</td>\n",
+       "      <td>-0.040817</td>\n",
+       "      <td>0.802806</td>\n",
+       "      <td>False</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "<p>5 rows × 47 columns</p>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "    DATE  STOCK  INDUSTRY  INDUSTRY_GROUP  SECTOR  SUB_INDUSTRY     RET_1  \\\n",
+       "ID                                                                          \n",
+       "0      0      2        18               5       3            44 -0.015748   \n",
+       "1      0      3        43              15       6           104  0.003984   \n",
+       "2      0      4        57              20       8           142  0.000440   \n",
+       "3      0      8         1               1       1             2  0.031298   \n",
+       "4      0     14        36              12       5            92  0.027273   \n",
+       "\n",
+       "    VOLUME_1     RET_2  VOLUME_2  ...  VOLUME_16    RET_17  VOLUME_17  \\\n",
+       "ID                                ...                                   \n",
+       "0   0.147931 -0.015504  0.179183  ...   0.630899  0.003254  -0.379412   \n",
+       "1        NaN -0.090580       NaN  ...        NaN  0.003774        NaN   \n",
+       "2  -0.096282 -0.058896  0.084771  ...  -0.010336 -0.017612  -0.354333   \n",
+       "3  -0.429540  0.007756 -0.089919  ...   0.012105  0.033824  -0.290178   \n",
+       "4  -0.847155 -0.039302 -0.943033  ...  -0.277083 -0.012659   0.139086   \n",
+       "\n",
+       "      RET_18  VOLUME_18    RET_19  VOLUME_19    RET_20  VOLUME_20    RET  \n",
+       "ID                                                                        \n",
+       "0   0.008752  -0.110597 -0.012959   0.174521 -0.002155  -0.000937   True  \n",
+       "1  -0.018518        NaN -0.028777        NaN -0.034722        NaN   True  \n",
+       "2  -0.006562  -0.519391 -0.012101  -0.356157 -0.006867  -0.308868  False  \n",
+       "3  -0.001468  -0.663834 -0.013520  -0.562126 -0.036745  -0.631458  False  \n",
+       "4   0.004237  -0.017547  0.004256   0.579510 -0.040817   0.802806  False  \n",
+       "\n",
+       "[5 rows x 47 columns]"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "x_train = pd.read_csv('./x_train.csv', index_col='ID')\n",
+    "y_train = pd.read_csv('./y_train.csv', index_col='ID')\n",
+    "train = pd.concat([x_train, y_train], axis=1)\n",
+    "test = pd.read_csv('./x_test.csv', index_col='ID')\n",
+    "train.head()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Feature Engineering\n",
+    "\n",
+    "The main drawback in this challenge would be to deal with the noise. To do that, we could create some feature that aggregate features with some statistics. \n",
+    "\n",
+    "The following cell computes statistics on a given target conditionally to some features. For example, we want to generate a feature that describe the mean of `RET_1` conditionally to the `SECTOR` and the `DATE`.\n",
+    "\n",
+    "**Ideas of improvement**: change shifts, the conditional features, the statistics, and the target. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Feature engineering\n",
+    "new_features = []\n",
+    "\n",
+    "# Conditional aggregated features\n",
+    "shifts = [1]  # Choose some different shifts\n",
+    "statistics = ['mean']  # the type of stat\n",
+    "gb_features = ['SECTOR', 'DATE']\n",
+    "target_feature = 'RET'\n",
+    "tmp_name = '_'.join(gb_features)\n",
+    "for shift in shifts:\n",
+    "    for stat in statistics:\n",
+    "        name = f'{target_feature}_{shift}_{tmp_name}_{stat}'\n",
+    "        feat = f'{target_feature}_{shift}'\n",
+    "        new_features.append(name)\n",
+    "        for data in [train, test]:\n",
+    "            data[name] = data.groupby(gb_features)[feat].transform(stat)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Feature selection\n",
+    "\n",
+    "To reduce the number of feature (and the noise) we only consider the 5 last days of `RET` and `VOLUME` in addition to the newly created feature."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>RET_1</th>\n",
+       "      <th>RET_2</th>\n",
+       "      <th>RET_3</th>\n",
+       "      <th>RET_4</th>\n",
+       "      <th>RET_5</th>\n",
+       "      <th>VOLUME_1</th>\n",
+       "      <th>VOLUME_2</th>\n",
+       "      <th>VOLUME_3</th>\n",
+       "      <th>VOLUME_4</th>\n",
+       "      <th>VOLUME_5</th>\n",
+       "      <th>RET_1_SECTOR_DATE_mean</th>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>ID</th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "      <th></th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>0</th>\n",
+       "      <td>-0.015748</td>\n",
+       "      <td>-0.015504</td>\n",
+       "      <td>0.010972</td>\n",
+       "      <td>-0.014672</td>\n",
+       "      <td>0.016483</td>\n",
+       "      <td>0.147931</td>\n",
+       "      <td>0.179183</td>\n",
+       "      <td>0.033832</td>\n",
+       "      <td>-0.362868</td>\n",
+       "      <td>-0.972920</td>\n",
+       "      <td>0.009178</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>1</th>\n",
+       "      <td>0.003984</td>\n",
+       "      <td>-0.090580</td>\n",
+       "      <td>0.018826</td>\n",
+       "      <td>-0.025540</td>\n",
+       "      <td>-0.038062</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>NaN</td>\n",
+       "      <td>0.006477</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>2</th>\n",
+       "      <td>0.000440</td>\n",
+       "      <td>-0.058896</td>\n",
+       "      <td>-0.009042</td>\n",
+       "      <td>0.024852</td>\n",
+       "      <td>0.009354</td>\n",
+       "      <td>-0.096282</td>\n",
+       "      <td>0.084771</td>\n",
+       "      <td>-0.298777</td>\n",
+       "      <td>-0.157421</td>\n",
+       "      <td>0.091455</td>\n",
+       "      <td>0.013449</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>3</th>\n",
+       "      <td>0.031298</td>\n",
+       "      <td>0.007756</td>\n",
+       "      <td>-0.004632</td>\n",
+       "      <td>-0.019677</td>\n",
+       "      <td>0.003544</td>\n",
+       "      <td>-0.429540</td>\n",
+       "      <td>-0.089919</td>\n",
+       "      <td>-0.639737</td>\n",
+       "      <td>-0.940163</td>\n",
+       "      <td>-0.882464</td>\n",
+       "      <td>0.017253</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>4</th>\n",
+       "      <td>0.027273</td>\n",
+       "      <td>-0.039302</td>\n",
+       "      <td>0.000000</td>\n",
+       "      <td>0.000000</td>\n",
+       "      <td>0.022321</td>\n",
+       "      <td>-0.847155</td>\n",
+       "      <td>-0.943033</td>\n",
+       "      <td>-1.180629</td>\n",
+       "      <td>-1.313896</td>\n",
+       "      <td>-1.204398</td>\n",
+       "      <td>0.006241</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "       RET_1     RET_2     RET_3     RET_4     RET_5  VOLUME_1  VOLUME_2  \\\n",
+       "ID                                                                         \n",
+       "0  -0.015748 -0.015504  0.010972 -0.014672  0.016483  0.147931  0.179183   \n",
+       "1   0.003984 -0.090580  0.018826 -0.025540 -0.038062       NaN       NaN   \n",
+       "2   0.000440 -0.058896 -0.009042  0.024852  0.009354 -0.096282  0.084771   \n",
+       "3   0.031298  0.007756 -0.004632 -0.019677  0.003544 -0.429540 -0.089919   \n",
+       "4   0.027273 -0.039302  0.000000  0.000000  0.022321 -0.847155 -0.943033   \n",
+       "\n",
+       "    VOLUME_3  VOLUME_4  VOLUME_5  RET_1_SECTOR_DATE_mean  \n",
+       "ID                                                        \n",
+       "0   0.033832 -0.362868 -0.972920                0.009178  \n",
+       "1        NaN       NaN       NaN                0.006477  \n",
+       "2  -0.298777 -0.157421  0.091455                0.013449  \n",
+       "3  -0.639737 -0.940163 -0.882464                0.017253  \n",
+       "4  -1.180629 -1.313896 -1.204398                0.006241  "
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "target = 'RET'\n",
+    "\n",
+    "n_shifts = 5  # If you don't want all the shifts to reduce noise\n",
+    "features = ['RET_%d' % (i + 1) for i in range(n_shifts)]\n",
+    "features += ['VOLUME_%d' % (i + 1) for i in range(n_shifts)]\n",
+    "features += new_features  # The conditional features\n",
+    "train[features].head()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Model and local score\n",
+    "\n",
+    "A Random Forest (RF) model is chosen for the Benchmark. We consider a large number of tree with a quiet small depth. The missing values are simply filled with 0. A KFold is done on the dates (using `DATE`) for a local scoring of the model. \n",
+    "\n",
+    "**Ideas of improvements**: Tune the RF hyperparameters, deal with the missing values, change the features, consider another model, ..."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Fold 1 - Accuracy: 51.87%\n",
+      "Fold 2 - Accuracy: 50.65%\n",
+      "Fold 3 - Accuracy: 51.11%\n",
+      "Fold 4 - Accuracy: 52.10%\n",
+      "Accuracy: 51.43% [50.85 ; 52.02] (+- 0.58)\n"
+     ]
+    }
+   ],
+   "source": [
+    "X_train = train[features]\n",
+    "y_train = train[target]\n",
+    "\n",
+    "# A quiet large number of trees with low depth to prevent overfits\n",
+    "rf_params = {\n",
+    "    'n_estimators': 500,\n",
+    "    'max_depth': 2**3,\n",
+    "    'random_state': 0,\n",
+    "    'n_jobs': -1\n",
+    "}\n",
+    "\n",
+    "train_dates = train['DATE'].unique()\n",
+    "test_dates = test['DATE'].unique()\n",
+    "\n",
+    "n_splits = 4\n",
+    "scores = []\n",
+    "models = []\n",
+    "\n",
+    "splits = KFold(n_splits=n_splits, random_state=0,\n",
+    "               shuffle=True).split(train_dates)\n",
+    "\n",
+    "for i, (local_train_dates_ids, local_test_dates_ids) in enumerate(splits):\n",
+    "    local_train_dates = train_dates[local_train_dates_ids]\n",
+    "    local_test_dates = train_dates[local_test_dates_ids]\n",
+    "\n",
+    "    local_train_ids = train['DATE'].isin(local_train_dates)\n",
+    "    local_test_ids = train['DATE'].isin(local_test_dates)\n",
+    "\n",
+    "    X_local_train = X_train.loc[local_train_ids]\n",
+    "    y_local_train = y_train.loc[local_train_ids]\n",
+    "    X_local_test = X_train.loc[local_test_ids]\n",
+    "    y_local_test = y_train.loc[local_test_ids]\n",
+    "\n",
+    "    X_local_train = X_local_train.fillna(0)\n",
+    "    X_local_test = X_local_test.fillna(0)\n",
+    "\n",
+    "    model = RandomForestClassifier(**rf_params)\n",
+    "    model.fit(X_local_train, y_local_train)\n",
+    "\n",
+    "    y_local_pred = model.predict_proba(X_local_test)[:, 1]\n",
+    "    \n",
+    "    sub = train.loc[local_test_ids].copy()\n",
+    "    sub['pred'] = y_local_pred\n",
+    "    y_local_pred = sub.groupby('DATE')['pred'].transform(lambda x: x > x.median()).values\n",
+    "\n",
+    "    models.append(model)\n",
+    "    score = accuracy_score(y_local_test, y_local_pred)\n",
+    "    scores.append(score)\n",
+    "    print(f\"Fold {i+1} - Accuracy: {score* 100:.2f}%\")\n",
+    "\n",
+    "mean = np.mean(scores)*100\n",
+    "std = np.std(scores)*100\n",
+    "u = (mean + std)\n",
+    "l = (mean - std)\n",
+    "print(f'Accuracy: {mean:.2f}% [{l:.2f} ; {u:.2f}] (+- {std:.2f})')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<matplotlib.axes._subplots.AxesSubplot at 0x1a384bc7848>"
+      ]
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
+    },
+    {
+     "data": {
+      "image/png": "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\n",
+      "text/plain": [
+       "<Figure size 432x288 with 1 Axes>"
+      ]
+     },
+     "metadata": {
+      "needs_background": "light"
+     },
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "feature_importances = pd.DataFrame([model.feature_importances_ for model in models], columns=features)\n",
+    "\n",
+    "sns.barplot(data=feature_importances, orient='h', order=feature_importances.mean().sort_values(ascending=False).index)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Generate the submission\n",
+    "\n",
+    "The same parameters of the RF model are considered. With that we build a new RF model on the entire `train` dataset. The predictions are saved in a `.csv` file."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "X_test = test[features]\n",
+    "\n",
+    "rf_params['random_state'] = 0\n",
+    "model = RandomForestClassifier(**rf_params)\n",
+    "model.fit(X_train.fillna(0), y_train)\n",
+    "y_pred = model.predict_proba(X_test.fillna(0))[:, 1]\n",
+    "\n",
+    "sub = test.copy()\n",
+    "sub['pred'] = y_pred\n",
+    "y_pred = sub.groupby('DATE')['pred'].transform(\n",
+    "    lambda x: x > x.median()).values\n",
+    "\n",
+    "submission = pd.Series(y_pred)\n",
+    "submission.index = test.index\n",
+    "submission.name = target\n",
+    "\n",
+    "submission.to_csv('./benchmark_qrt.csv', index=True, header=True)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "\n",
+    "The local accuracy is around 51. If we did not overfit, we shall expect something within the range above.\n",
+    "\n",
+    "After submitting the benchmark file at https://challengedata.ens.fr, we obtain a public score of 51.31 %."
+   ]
+  }
+ ],
+ "metadata": {
+  "hide_input": false,
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.7.5"
+  },
+  "toc": {
+   "base_numbering": 1,
+   "nav_menu": {},
+   "number_sections": true,
+   "sideBar": true,
+   "skip_h1_title": false,
+   "title_cell": "Table of Contents",
+   "title_sidebar": "Contents",
+   "toc_cell": false,
+   "toc_position": {},
+   "toc_section_display": true,
+   "toc_window_display": false
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}