diff options
| -rw-r--r-- | python/notebooks/dispersion_tranche_model.ipynb | 320 |
1 files changed, 169 insertions, 151 deletions
diff --git a/python/notebooks/dispersion_tranche_model.ipynb b/python/notebooks/dispersion_tranche_model.ipynb index a56332ac..a2bb5a67 100644 --- a/python/notebooks/dispersion_tranche_model.ipynb +++ b/python/notebooks/dispersion_tranche_model.ipynb @@ -13,15 +13,13 @@ "import exploration.dispersion as disp\n", "import matplotlib.pyplot as plt\n", "import statsmodels.formula.api as smf\n", - "import analytics.tranche_data as tdata\n", - "import ipysheet\n", + "import serenitas.analytics.tranche_data as tdata\n", "\n", - "from analytics.basket_index import MarkitBasketIndex\n", - "from analytics import on_the_run\n", + "from serenitas.analytics.basket_index import MarkitBasketIndex\n", + "from serenitas.analytics import on_the_run\n", "from statsmodels.graphics.regressionplots import plot_fit\n", "from scipy.special import logit, expit\n", - "from pygam import LinearGAM, s, f, GAM\n", - "from utils.db import dbengine, dbconn" + "from serenitas.utils.db import dbengine, dbconn" ] }, { @@ -30,89 +28,13 @@ "metadata": {}, "outputs": [], "source": [ - "%matplotlib inline" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "value_date = (datetime.datetime.today() - pd.offsets.BDay(1)).date()\n", - "start = (datetime.datetime.today() - pd.offsets.BDay(1) * 365 *4).date()\n", - "#end = (start + pd.offsets.BDay(1) * 365).date()\n", - "end = datetime.datetime.today()\n", - "gini_model, gini_results = {}, {}\n", - "conn = dbconn(\"serenitasdb\")\n", - "conn.autocommit = True\n", - "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " risk = disp.get_tranche_data(dbconn(\"serenitasdb\"), index_type)\n", - " risk = risk[risk.index_duration > 1] #filter out the short duration ones\n", - " gini_results[index_type], gini_model[index_type] = disp.create_models_v2(conn, risk)\n", - " #fitted = gini_model[index_type].fit()\n", - " #w = 1/(expit(fitted.fittedvalues + fitted.resid) -expit(fitted.fittedvalues))**2\n", - " #gini_results[index_type], gini_model[index_type] = disp.create_models_v2(conn, risk, w)\n", - "gini_model['HY'].fit().summary()\n", - "\n", - "fieldlist = ['exp_percentage','dispersion','gini','tranche_loss_per','mispricing']\n", - "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/' + index_type + '_results.csv')" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "fieldlist = ['exp_percentage','dispersion','gini','tranche_loss_per','mispricing']\n", - "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/' + index_type + '_results.csv')" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "#plot the residuals\n", - "fitted = gini_model['HY'].fit()\n", - "plt.figure(figsize=(8,5))\n", - "p=plt.scatter(x=expit(fitted.fittedvalues),y=expit(fitted.fittedvalues + fitted.resid) -expit(fitted.fittedvalues),edgecolor='k')\n", - "xmin=min(expit(fitted.fittedvalues))\n", - "xmax = max(expit(fitted.fittedvalues))\n", - "plt.hlines(y=0,xmin=xmin*0.9,xmax=xmax*1.1,color='red',linestyle='--',lw=3)\n", - "plt.xlabel(\"Fitted values\",fontsize=15)\n", - "plt.ylabel(\"Residuals\",fontsize=15)\n", - "plt.title(\"Fitted vs. residuals plot\",fontsize=18)\n", - "plt.grid(True)\n", - "plt.show()" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "#plot the gini coefficients\n", - "ginis = gini_results['HY'].xs([0, '5yr', 'HY'],level=['attach','tenor', 'index']).groupby(['date', 'series']).nth(-1).gini.unstack(level='series')\n", - "ginis.plot()" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": {}, - "outputs": [], - "source": [ - "#sheet = ipysheet.sheet(rows=2000, columns=7, column_headers=False, row_headers=False)\n", - "import IPython\n", - "pd.set_option(\"display.max_rows\", None)\n", - "#IPython.OutputArea.auto_scroll_threshold = 20\n", - "ginis.sort_index(ascending=False)" + "from sklearn.preprocessing import PolynomialFeatures, PowerTransformer\n", + "from sklearn.linear_model import LinearRegression\n", + "from sklearn.pipeline import make_pipeline\n", + "from sklearn.feature_selection import RFECV, RFE\n", + "from sklearn.compose import TransformedTargetRegressor\n", + "from sklearn.model_selection import train_test_split\n", + "from sklearn.ensemble import RandomForestRegressor, GradientBoostingRegressor, VotingRegressor" ] }, { @@ -121,7 +43,7 @@ "metadata": {}, "outputs": [], "source": [ - "#look at Volatility vs Correlation" + "%matplotlib inline" ] }, { @@ -130,25 +52,26 @@ "metadata": {}, "outputs": [], "source": [ - "#use RFE to get the model instead\n", - "from sklearn.preprocessing import PolynomialFeatures, PowerTransformer, normalize\n", - "from sklearn.linear_model import LinearRegression\n", - "from sklearn.pipeline import make_pipeline\n", - "from sklearn.feature_selection import RFECV, SelectKBest, f_regression\n", - "from sklearn.compose import TransformedTargetRegressor\n", - "\n", - "import numpy as np\n", - "\n", - "class MyTransformedTargetRegressor(TransformedTargetRegressor):\n", - " @property\n", - " def coef_(self):\n", - " return self.regressor_.coef_\n", - " \n", - " @property\n", - " def feature_importances_(self):\n", - " return self.regressor_.feature_importances_\n", - " \n", - "index_type = 'HY'" + "#Section 1----------------------------------------------------\n", + "#RFE\n", + "drop_variable_list = ['tranche_loss_per', 'tranche_id', 'index_price', 'detach', 'corr_at_detach', \n", + " 'corr01', 'exp_percentage', 'indexfactor', 'duration', 'index_expected_loss',\n", + " 'index_theta', 'delta', 'expected_loss', 'attach_adj', 'detach_adj',\n", + " 'theta', 'cumulativeloss', \n", + " 'forward_delta', \n", + " #Comment out to include\n", + " 'index_duration',\n", + " # 'thickness',\n", + " # 'moneyness',\n", + " # 'index_basis',\n", + " 'att_moneyness', \n", + " 'det_moneyness',\n", + " 'dispersion',\n", + " # 'gini', \n", + " 'gamma',\n", + " 'theta',\n", + " 'index_theta'\n", + " ]" ] }, { @@ -157,21 +80,6 @@ "metadata": {}, "outputs": [], "source": [ - "all_variable_list = ['tranche_loss_per', 'tranche_id', 'index_price', 'index_basis', 'detach', 'corr_at_detach', \n", - " 'attach_adj', 'detach_adj', 'index_theta', 'delta','gamma', 'corr01', 'expected_loss', \n", - " 'exp_percentage', 'indexfactor', 'duration', 'index_expected_loss',\n", - " #Comment out to include\n", - " # 'index_duration',\n", - " 'theta',\n", - " 'cumulativeloss',\n", - " # 'att_moneyness',\n", - " 'det_moneyness',\n", - " 'thickness',\n", - " # 'moneyness',\n", - " # 'dispersion',\n", - " 'gini',\n", - " 'forward_delta']\n", - "\n", "def run_rfe(index_type):\n", " risk = disp.get_tranche_data(dbconn(\"serenitasdb\"), index_type)\n", " attach_max = risk.index.get_level_values(\"attach\").max()\n", @@ -179,18 +87,44 @@ " bottom_stack = bottom_stack[bottom_stack.tranche_loss_per > 0].dropna()\n", "\n", " #prepare the variables\n", - " y = bottom_stack['tranche_loss_per']\n", - " X = bottom_stack.drop(all_variable_list, axis=1)\n", - "\n", - " poly = PolynomialFeatures(3)\n", - " X = pd.DataFrame(PowerTransformer().fit_transform(X), index=X.index, columns=X.columns)\n", - " X_p = pd.DataFrame(poly.fit_transform(X), columns= poly.get_feature_names(X.columns))\n", - " regr = MyTransformedTargetRegressor(regressor=LinearRegression(), func=logit, inverse_func=expit)\n", - "\n", - " rfecv = RFECV(regr).fit(X_p,y)\n", + " y = logit(bottom_stack['tranche_loss_per'])\n", + " X = bottom_stack.drop(drop_variable_list, axis=1)\n", + " \n", + " X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)\n", + " \n", + " pipe_rfe = make_pipeline (PowerTransformer(),\n", + " PolynomialFeatures(interaction_only=True),\n", + " RFECV(estimator=LinearRegression(), \n", + " cv=10,\n", + " min_features_to_select=1))\n", + " \n", + " pipe_rfe.fit(X_train, y_train)\n", + " n_features_to_select = pipe_rfe['rfecv'].n_features_\n", + " pipe_rfe.steps[-1]= ('rfe', RFE(estimator=LinearRegression(), n_features_to_select = n_features_to_select))\n", + " model = pipe_rfe.fit(X_train, y_train)\n", + " \n", + " #RandomForest\n", + " #params = {'n_estimators': 100,\n", + " # 'min_samples_split': 3,\n", + " # 'verbose':1,\n", + " # 'n_jobs': -1}\n", + " #randomforest = RandomForestRegressor(**params)\n", + " \n", + " \n", + " #gradientboost\n", + " #params = {'n_estimators': 500,\n", + " # 'max_depth': 10,\n", + " # 'min_samples_split': 3,\n", + " # 'learning_rate': 0.01,\n", + " # 'loss': 'huber',\n", + " # 'verbose':1}\n", + " #gb = GradientBoostingRegressor(**params).fit(X_p, y)\n", + " \n", + " #model = VotingRegressor([('rf', model), ('lr', randomforest)]).fit(X_train, y_train)\n", + " #model = VotingRegressor([('lr', pipe_rfe)]).fit(X, logit(y))\n", "\n", " df = pd.merge(risk, \n", - " pd.DataFrame(rfecv.predict(X_p), \n", + " pd.DataFrame(expit(model.predict(X)), \n", " index=X.index, \n", " columns=['predict_tranche_loss']),\n", " how='left', left_index=True, right_index=True)\n", @@ -210,18 +144,19 @@ " * df.index_expected_loss\n", " / (df.detach_adj - df.attach_adj)\n", " )\n", - "\n", - " print(index_type, \" Chosen columns: \", X_p[X_p.columns[rfecv.support_]].columns)\n", - " print(index_type, \" Score: \", rfecv.score(X_p, y))\n", + " rfe_result = pipe_rfe\n", + " print(index_type, \" num features: \", n_features_to_select)\n", + " print(index_type, \" Chosen columns: \", np.array(rfe_result['polynomialfeatures'].get_feature_names(X.columns))[rfe_result['rfe'].support_])\n", + " print(index_type, \" Training Score: \", model.score(X_train, y_train))\n", + " print(index_type, \" Testing Score: \", model.score(X_test, y_test))\n", " \n", - " return rfecv, df\n", + " return model, df, X\n", "\n", - "gini_model, gini_results = {}, {}\n", - "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_model[index_type], gini_results[index_type] = run_rfe(index_type)\n", + "gini_model, gini_results, gini_X = {}, {}, {}\n", "fieldlist = ['exp_percentage','dispersion','gini','tranche_loss_per','mispricing']\n", "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/' + index_type + '_results_rfecv.csv')" + " gini_model[index_type], gini_results[index_type], gini_X[index_type] = run_rfe(index_type)\n", + " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/DispersionModel/' + index_type + '_results_rfecv.csv')" ] }, { @@ -230,12 +165,70 @@ "metadata": {}, "outputs": [], "source": [ + "#examine the effect of any paricular variable\n", + "steps = 100\n", + "index_type = 'HY'\n", + "plots = {}\n", + "tranche_attach = []\n", + "\n", + "for i, X in gini_X[index_type].groupby('attach'):\n", + " tranche_attach.append(X.index[0][5])\n", + " for var in X.columns:\n", + " bins = np.linspace(X[var].min(), X[var].max(),num=steps)\n", + " testing_df = pd.DataFrame(bins, columns=[var])\n", + " for var_1 in X.drop(var, axis=1).columns:\n", + " testing_df = pd.concat([testing_df, pd.Series(np.repeat(X.iloc[-1][var_1], steps),name=var_1)], axis=1)\n", + " plots[i, var] = pd.Series(expit(gini_model[index_type].predict(testing_df[X.columns])), index=testing_df[var])\n", + "\n", + "fig, axes = plt.subplots(nrows=3, ncols=len(X.columns), figsize = (20,10))\n", + "for i, p in enumerate(plots):\n", + " x_loc = int(i/len(X.columns))\n", + " y_loc = i % len(X.columns)\n", + " if x_loc == 0:\n", + " axes[x_loc, y_loc].set_title(p[1]) \n", + " plots[p].plot(ax=axes[x_loc, y_loc], label=i, xlabel=\"\")\n", + "for i in [0,1,2]:\n", + " fig.axes[i*len(X.columns)].text(-0.2, 0.5, \"tranche attach: \" + str(tranche_attach[i]),\n", + " transform=fig.axes[i*len(X.columns)].transAxes,\n", + " verticalalignment='center',\n", + " rotation=90)\n", + "fig.savefig(\"/home/serenitas/edwin/PythonGraphs/dispersion_model.png\", bbox_inches='tight')\n", + "\n", + "fig_1, axes_1 = plt.subplots(nrows=3, ncols=1, figsize = (15,8))\n", + "for i, p in enumerate(plots):\n", + " x_loc = int(i/len(X.columns))\n", + " plots[p].plot(ax=axes_1[x_loc], label=p[1], xlabel=\"\", legend=True)\n", + "for i in [0,1,2]:\n", + " fig_1.axes[i].text(-0.05, 0.5, \"tranche attach: \" + str(tranche_attach[i]),\n", + " transform=fig_1.axes[i].transAxes,\n", + " verticalalignment='center',\n", + " rotation=90)\n", + "fig_1.savefig(\"/home/serenitas/edwin/PythonGraphs/dispersion_model_consolidated.png\", bbox_inches='tight')" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "#Section 2----------------------------------------------------\n", + "#Fixed Model\n", + "value_date = (datetime.datetime.today() - pd.offsets.BDay(1)).date()\n", + "start = (datetime.datetime.today() - pd.offsets.BDay(1) * 365 *4).date()\n", + "end = datetime.datetime.today()\n", "gini_model, gini_results = {}, {}\n", + "conn = dbconn(\"serenitasdb\")\n", + "conn.autocommit = True\n", "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_model[index_type], gini_results[index_type] = run_rfe(index_type)\n", + " risk = disp.get_tranche_data(dbconn(\"serenitasdb\"), index_type)\n", + " risk = risk[risk.index_duration > 1] #filter out the short duration ones\n", + " gini_results[index_type], gini_model[index_type] = disp.create_models_v2(conn, risk)\n", + "gini_model['HY'].fit().summary()\n", + "\n", "fieldlist = ['exp_percentage','dispersion','gini','tranche_loss_per','mispricing']\n", "for index_type in ['HY', 'IG', 'EU', 'XO']:\n", - " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/' + index_type + '_results_rfecv.csv')" + " gini_results[index_type][fieldlist].to_csv('/home/serenitas/edwin/DispersionModel/' + index_type + '_results.csv')" ] }, { @@ -244,14 +237,39 @@ "metadata": {}, "outputs": [], "source": [ - "# Plot number of features VS. cross-validation scores\n", - "index_type = 'IG'\n", - "plt.figure()\n", - "plt.xlabel(\"Number of features selected\")\n", - "plt.ylabel(\"Cross validation score (nb of correct classifications)\")\n", - "plt.plot(range(1, len(gini_model[index_type].grid_scores_) + 1), gini_model[index_type].grid_scores_)\n", + "#plot the residuals\n", + "fitted = gini_model['HY'].fit()\n", + "plt.figure(figsize=(8,5))\n", + "p=plt.scatter(x=expit(fitted.fittedvalues),y=expit(fitted.fittedvalues + fitted.resid) -expit(fitted.fittedvalues),edgecolor='k')\n", + "xmin=min(expit(fitted.fittedvalues))\n", + "xmax = max(expit(fitted.fittedvalues))\n", + "plt.hlines(y=0,xmin=xmin*0.9,xmax=xmax*1.1,color='red',linestyle='--',lw=3)\n", + "plt.xlabel(\"Fitted values\",fontsize=15)\n", + "plt.ylabel(\"Residuals\",fontsize=15)\n", + "plt.title(\"Fitted vs. residuals plot\",fontsize=18)\n", + "plt.grid(True)\n", "plt.show()" ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [ + "#Section 3----------------------------------------------------\n", + "#plot the gini coefficients\n", + "index_type='EU'\n", + "ginis = gini_results[index_type].xs([0, '5yr', index_type],level=['attach','tenor', 'index']).groupby(['date', 'series']).nth(-1).gini.unstack(level='series')\n", + "ginis.sort_index(ascending=False).to_clipboard()" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": {}, + "outputs": [], + "source": [] } ], "metadata": { @@ -270,7 +288,7 @@ "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", - "version": "3.8.5" + "version": "3.9.1-final" } }, "nbformat": 4, |