def model(base, winequality_vd): base.cursor.execute("DROP MODEL IF EXISTS linreg_model_test") model_class = LinearRegression("linreg_model_test", cursor=base.cursor) model_class.fit("public.winequality", ["citric_acid", "residual_sugar", "alcohol"], "quality") yield model_class model_class.drop()
def model(winequality_vd): model_class = LinearRegression("linreg_model_test", ) model_class.drop() model_class.fit("public.winequality", ["citric_acid", "residual_sugar", "alcohol"], "quality") yield model_class model_class.drop()
def test_contour(self, winequality_vd): model_test = LinearRegression("model_contour", ) model_test.drop() model_test.fit( winequality_vd, ["citric_acid", "residual_sugar"], "quality", ) result = model_test.contour() assert len(result.get_default_bbox_extra_artists()) == 32 model_test.drop()
def test_model_from_vDF(self, base, winequality_vd): base.cursor.execute("DROP MODEL IF EXISTS linreg_from_vDF") model_test = LinearRegression("linreg_from_vDF", cursor=base.cursor) model_test.fit(winequality_vd, ["alcohol"], "quality") base.cursor.execute( "SELECT model_name FROM models WHERE model_name = 'linreg_from_vDF'" ) assert base.cursor.fetchone()[0] == "linreg_from_vDF" model_test.drop()
def test_set_cursor(self, base): model_test = LinearRegression("linear_reg_cursor_test", cursor=base.cursor) # TODO: creat a new cursor model_test.set_cursor(base.cursor) model_test.drop() model_test.fit("public.winequality", ["alcohol"], "quality") base.cursor.execute( "SELECT model_name FROM models WHERE model_name = 'linear_reg_cursor_test'" ) assert base.cursor.fetchone()[0] == "linear_reg_cursor_test" model_test.drop()
def test_get_plot(self, winequality_vd): current_cursor().execute("DROP MODEL IF EXISTS model_test_plot") model_test = LinearRegression("model_test_plot", ) model_test.fit(winequality_vd, ["alcohol"], "quality") result = model_test.plot(color="r") assert len(result.get_default_bbox_extra_artists()) == 9 plt.close("all") model_test.drop() model_test.fit(winequality_vd, ["alcohol", "residual_sugar"], "quality") result = model_test.plot(color="r") assert len(result.get_default_bbox_extra_artists()) == 3 plt.close("all") model_test.drop()
def test_drop(self): current_cursor().execute("DROP MODEL IF EXISTS linreg_model_test_drop") model_test = LinearRegression("linreg_model_test_drop", ) model_test.fit("public.winequality", ["alcohol"], "quality") current_cursor().execute( "SELECT model_name FROM models WHERE model_name = 'linreg_model_test_drop'" ) assert current_cursor().fetchone()[0] == "linreg_model_test_drop" model_test.drop() current_cursor().execute( "SELECT model_name FROM models WHERE model_name = 'linreg_model_test_drop'" ) assert current_cursor().fetchone() is None
def test_drop(self, base): base.cursor.execute("DROP MODEL IF EXISTS linreg_model_test_drop") model_test = LinearRegression("linreg_model_test_drop", cursor=base.cursor) model_test.fit("public.winequality", ["alcohol"], "quality") base.cursor.execute( "SELECT model_name FROM models WHERE model_name = 'linreg_model_test_drop'" ) assert base.cursor.fetchone()[0] == "linreg_model_test_drop" model_test.drop() base.cursor.execute( "SELECT model_name FROM models WHERE model_name = 'linreg_model_test_drop'" ) assert base.cursor.fetchone() is None
def test_cochrane_orcutt(self, airline_vd): airline_copy = airline_vd.copy() airline_copy["passengers_bias"] = (airline_copy["passengers"]**2 - 50 * st.random()) drop("lin_cochrane_orcutt_model_test", method="model") model = LinearRegression("lin_cochrane_orcutt_model_test") model.fit(airline_copy, ["passengers_bias"], "passengers") result = st.cochrane_orcutt( model, airline_copy, ts="date", prais_winsten=True, ) assert result.coef_["coefficient"][0] == pytest.approx( 25.8582027191416, 1e-2) assert result.coef_["coefficient"][1] == pytest.approx( 0.00123563974547625, 1e-2) model.drop()
def test_cochrane_orcutt(self, airline_vd, base): airline_copy = airline_vd.copy() airline_copy["passengers_bias"] = airline_copy[ "passengers"]**2 - 50 * st.random() from verticapy.learn.linear_model import LinearRegression base.cursor.execute( "DROP MODEL IF EXISTS lin_cochrane_orcutt_model_test") model = LinearRegression("lin_cochrane_orcutt_model_test", cursor=base.cursor) model.fit(airline_copy, ["passengers_bias"], "passengers") result = st.cochrane_orcutt( model, airline_copy, ts="date", prais_winsten=True, ) assert result.coef_["coefficient"][0] == pytest.approx( 25.8582027191416, 1e-2) assert result.coef_["coefficient"][1] == pytest.approx( 0.00123563974547625, 1e-2) model.drop()
def het_white( vdf: vDataFrame, eps: str, X: list, ): """ --------------------------------------------------------------------------- White’s Lagrange Multiplier Test for heteroscedasticity. Parameters ---------- vdf: vDataFrame Input vDataFrame. eps: str Input residual vcolumn. X: str Exogenous Variables to test the heteroscedasticity on. Returns ------- tablesample An object containing the result. For more information, see utilities.tablesample. """ check_types( [("eps", eps, [str],), ("X", X, [list],), ("vdf", vdf, [vDataFrame, str,],),], ) columns_check([eps] + X, vdf) eps = vdf_columns_names([eps], vdf)[0] X = vdf_columns_names(X, vdf) X_0 = ["1"] + X variables = [] variables_names = [] for i in range(len(X_0)): for j in range(i, len(X_0)): if i != 0 or j != 0: variables += ["{} * {} AS var_{}_{}".format(X_0[i], X_0[j], i, j)] variables_names += ["var_{}_{}".format(i, j)] query = "(SELECT {}, POWER({}, 2) AS VERTICAPY_TEMP_eps2 FROM {}) VERTICAPY_SUBTABLE".format( ", ".join(variables), eps, vdf.__genSQL__() ) vdf_white = vdf_from_relation(query, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) from verticapy.learn.linear_model import LinearRegression schema_writing = vdf._VERTICAPY_VARIABLES_["schema_writing"] if not (schema_writing): schema_writing = "public" name = schema_writing + ".VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format( get_session(vdf._VERTICAPY_VARIABLES_["cursor"]) ) model = LinearRegression(name, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) try: model.fit(vdf_white, variables_names, "VERTICAPY_TEMP_eps2") R2 = model.score("r2") model.drop() except: try: model.set_params({"solver": "bfgs"}) model.fit(vdf_white, variables_names, "VERTICAPY_TEMP_eps2") R2 = model.score("r2") model.drop() except: model.drop() raise n = vdf.shape()[0] if len(X) > 1: k = 2 * len(X) + math.factorial(len(X)) / 2 / (math.factorial(len(X) - 2)) else: k = 1 LM = n * R2 lm_pvalue = chi2.sf(LM, k) F = (n - k - 1) * R2 / (1 - R2) / k f_pvalue = f.sf(F, k, n - k - 1) result = tablesample( { "index": [ "Lagrange Multiplier Statistic", "lm_p_value", "F Value", "f_p_value", ], "value": [LM, lm_pvalue, F, f_pvalue], } ) return result
def het_goldfeldquandt( vdf: vDataFrame, y: str, X: list, idx: int = 0, split: float = 0.5 ): """ --------------------------------------------------------------------------- Goldfeld-Quandt homoscedasticity test. Parameters ---------- vdf: vDataFrame Input vDataFrame. y: str Response Column. X: list Exogenous Variables. idx: int, optional Column index of variable according to which observations are sorted for the split. split: float, optional Float to indicate where to split (Example: 0.5 to split on the median). Returns ------- tablesample An object containing the result. For more information, see utilities.tablesample. """ def model_fit(input_relation, X, y, model): var = [] for vdf_tmp in input_relation: model.drop() model.fit(vdf_tmp, X, y) model.predict(vdf_tmp, name="verticapy_prediction") vdf_tmp["residual_0"] = vdf_tmp[y] - vdf_tmp["verticapy_prediction"] var += [vdf_tmp["residual_0"].var()] model.drop() return var check_types( [ ("y", y, [str],), ("X", X, [list],), ("idx", idx, [int, float],), ("split", split, [int, float],), ("vdf", vdf, [vDataFrame, str,],), ], ) columns_check([y] + X, vdf) y = vdf_columns_names([y], vdf)[0] X = vdf_columns_names(X, vdf) split_value = vdf[X[idx]].quantile(split) vdf_0_half = vdf.search(vdf[X[idx]] < split_value) vdf_1_half = vdf.search(vdf[X[idx]] > split_value) from verticapy.learn.linear_model import LinearRegression schema_writing = vdf._VERTICAPY_VARIABLES_["schema_writing"] if not (schema_writing): schema_writing = "public" name = schema_writing + ".VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format( get_session(vdf._VERTICAPY_VARIABLES_["cursor"]) ) model = LinearRegression(name, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) try: var0, var1 = model_fit([vdf_0_half, vdf_1_half], X, y, model) except: try: model.set_params({"solver": "bfgs"}) var0, var1 = model_fit([vdf_0_half, vdf_1_half], X, y, model) except: model.drop() raise n, m = vdf_0_half.shape()[0], vdf_1_half.shape()[0] F = var0 / var1 f_pvalue = f.sf(F, n, m) result = tablesample({"index": ["F Value", "f_p_value",], "value": [F, f_pvalue],}) return result
def het_breuschpagan( vdf: vDataFrame, eps: str, X: list, ): """ --------------------------------------------------------------------------- Breusch-Pagan test for heteroscedasticity. Parameters ---------- vdf: vDataFrame Input vDataFrame. eps: str Input residual vcolumn. X: list Exogenous Variables to test the heteroscedasticity on. Returns ------- tablesample An object containing the result. For more information, see utilities.tablesample. """ check_types( [("eps", eps, [str],), ("X", X, [list],), ("vdf", vdf, [vDataFrame, str,],),], ) columns_check([eps] + X, vdf) eps = vdf_columns_names([eps], vdf)[0] X = vdf_columns_names(X, vdf) from verticapy.learn.linear_model import LinearRegression schema_writing = vdf._VERTICAPY_VARIABLES_["schema_writing"] if not (schema_writing): schema_writing = "public" name = schema_writing + ".VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format( get_session(vdf._VERTICAPY_VARIABLES_["cursor"]) ) model = LinearRegression(name, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) vdf_copy = vdf.copy() vdf_copy["VERTICAPY_TEMP_eps2"] = vdf_copy[eps] ** 2 try: model.fit(vdf_copy, X, "VERTICAPY_TEMP_eps2") R2 = model.score("r2") model.drop() except: try: model.set_params({"solver": "bfgs"}) model.fit(vdf_copy, X, "VERTICAPY_TEMP_eps2") R2 = model.score("r2") model.drop() except: model.drop() raise n = vdf.shape()[0] k = len(X) LM = n * R2 lm_pvalue = chi2.sf(LM, k) F = (n - k - 1) * R2 / (1 - R2) / k f_pvalue = f.sf(F, k, n - k - 1) result = tablesample( { "index": [ "Lagrange Multiplier Statistic", "lm_p_value", "F Value", "f_p_value", ], "value": [LM, lm_pvalue, F, f_pvalue], } ) return result
def het_arch( vdf: vDataFrame, eps: str, ts: str, by: list = [], p: int = 1, ): """ --------------------------------------------------------------------------- Engle’s Test for Autoregressive Conditional Heteroscedasticity (ARCH). Parameters ---------- vdf: vDataFrame Input vDataFrame. eps: str Input residual vcolumn. ts: str vcolumn used as timeline. It will be to use to order the data. It can be a numerical or type date like (date, datetime, timestamp...) vcolumn. by: list, optional vcolumns used in the partition. p: int, optional Number of lags to consider in the test. Returns ------- tablesample An object containing the result. For more information, see utilities.tablesample. """ check_types( [ ("eps", eps, [str],), ("ts", ts, [str],), ("p", p, [int, float],), ("vdf", vdf, [vDataFrame, str,],), ], ) columns_check([eps, ts] + by, vdf) eps = vdf_columns_names([eps], vdf)[0] ts = vdf_columns_names([ts], vdf)[0] by = vdf_columns_names(by, vdf) X = [] X_names = [] for i in range(0, p + 1): X += [ "LAG(POWER({}, 2), {}) OVER({}ORDER BY {}) AS lag_{}".format( eps, i, ("PARTITION BY " + ", ".join(by)) if (by) else "", ts, i ) ] X_names += ["lag_{}".format(i)] query = "(SELECT {} FROM {}) VERTICAPY_SUBTABLE".format( ", ".join(X), vdf.__genSQL__() ) vdf_lags = vdf_from_relation(query, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) from verticapy.learn.linear_model import LinearRegression schema_writing = vdf._VERTICAPY_VARIABLES_["schema_writing"] if not (schema_writing): schema_writing = "public" name = schema_writing + ".VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format( get_session(vdf._VERTICAPY_VARIABLES_["cursor"]) ) model = LinearRegression(name, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) try: model.fit(vdf_lags, X_names[1:], X_names[0]) R2 = model.score("r2") model.drop() except: try: model.set_params({"solver": "bfgs"}) model.fit(vdf_lags, X_names[1:], X_names[0]) R2 = model.score("r2") model.drop() except: model.drop() raise n = vdf.shape()[0] k = len(X) LM = (n - p) * R2 lm_pvalue = chi2.sf(LM, p) F = (n - 2 * p - 1) * R2 / (1 - R2) / p f_pvalue = f.sf(F, p, n - 2 * p - 1) result = tablesample( { "index": [ "Lagrange Multiplier Statistic", "lm_p_value", "F Value", "f_p_value", ], "value": [LM, lm_pvalue, F, f_pvalue], } ) return result
def variance_inflation_factor( vdf: vDataFrame, X: list, X_idx: int = None, ): """ --------------------------------------------------------------------------- Computes the variance inflation factor (VIF). It can be used to detect multicollinearity in an OLS Regression Analysis. Parameters ---------- vdf: vDataFrame Input vDataFrame. X: list Input Variables. X_idx: int Index of the exogenous variable in X. If left to None, a tablesample will be returned with all the variables VIF. Returns ------- float VIF. """ check_types( [ ("X_idx", X_idx, [int],), ("X", X, [list],), ("vdf", vdf, [vDataFrame, str,],), ], ) columns_check(X, vdf) X = vdf_columns_names(X, vdf) if isinstance(X_idx, str): columns_check([X_idx], vdf) for i in range(len(X)): if str_column(X[i]) == str_column(X_idx): X_idx = i break if isinstance(X_idx, (int, float)): X_r = [] for i in range(len(X)): if i != X_idx: X_r += [X[i]] y_r = X[X_idx] from verticapy.learn.linear_model import LinearRegression schema_writing = vdf._VERTICAPY_VARIABLES_["schema_writing"] if not (schema_writing): schema_writing = "public" name = schema_writing + ".VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format( get_session(vdf._VERTICAPY_VARIABLES_["cursor"]) ) model = LinearRegression(name, cursor=vdf._VERTICAPY_VARIABLES_["cursor"]) try: model.fit(vdf, X_r, y_r) R2 = model.score("r2") model.drop() except: try: model.set_params({"solver": "bfgs"}) model.fit(vdf, X_r, y_r) R2 = model.score("r2") model.drop() except: model.drop() raise if 1 - R2 != 0: return 1 / (1 - R2) else: return np.inf elif X_idx == None: VIF = [] for i in range(len(X)): VIF += [variance_inflation_factor(vdf, X, i)] return tablesample({"X_idx": X, "VIF": VIF}) else: raise ParameterError( f"Wrong type for Parameter X_idx.\nExpected integer, found {type(X_idx)}." )
def seasonal_decompose( vdf: vDataFrame, column: str, ts: str, by: list = [], period: int = -1, polynomial_order: int = 1, estimate_seasonality: bool = True, rule: Union[str, datetime.timedelta] = None, mult: bool = False, two_sided: bool = False, ): """ --------------------------------------------------------------------------- Performs a seasonal time series decomposition. Parameters ---------- vdf: vDataFrame Input vDataFrame. column: str Input vcolumn to decompose. ts: str TS (Time Series) vcolumn to use to order the data. It can be of type date or a numerical vcolumn. by: list, optional vcolumns used in the partition. period: int, optional Time Series period. It is used to retrieve the seasonality component. if period <= 0, the seasonal component will be estimated using ACF. In this case, polynomial_order must be greater than 0. polynomial_order: int, optional If greater than 0, the trend will be estimated using a polynomial of degree 'polynomial_order'. The parameter 'two_sided' will be ignored. If equal to 0, the trend will be estimated using Moving Averages. estimate_seasonality: bool, optional If set to True, the seasonality will be estimated using cosine and sine functions. rule: str / time, optional Interval to use to slice the time. For example, '5 minutes' will create records separated by '5 minutes' time interval. mult: bool, optional If set to True, the decomposition type will be 'multiplicative'. Otherwise, it is 'additive'. two_sided: bool, optional If set to True, a centered moving average is used for the trend isolation. Otherwise only past values are used. Returns ------- vDataFrame object containing (ts, column, TS seasonal part, TS trend, TS noise). """ if isinstance(by, str): by = [by] check_types( [ ("ts", ts, [str],), ("column", column, [str],), ("by", by, [list],), ("rule", rule, [str, datetime.timedelta,],), ("vdf", vdf, [vDataFrame,],), ("period", period, [int,],), ("mult", mult, [bool,],), ("two_sided", two_sided, [bool,],), ("polynomial_order", polynomial_order, [int,],), ("estimate_seasonality", estimate_seasonality, [bool,],), ], ) assert period > 0 or polynomial_order > 0, ParameterError("Parameters 'polynomial_order' and 'period' can not be both null.") columns_check([column, ts] + by, vdf) ts, column, by = ( vdf_columns_names([ts], vdf)[0], vdf_columns_names([column], vdf)[0], vdf_columns_names(by, vdf), ) if rule: vdf_tmp = vdf.asfreq(ts=ts, rule=period, method={column: "linear"}, by=by) else: vdf_tmp = vdf[[ts, column]] trend_name, seasonal_name, epsilon_name = ( "{}_trend".format(column[1:-1]), "{}_seasonal".format(column[1:-1]), "{}_epsilon".format(column[1:-1]), ) by, by_tmp = "" if not (by) else "PARTITION BY " + ", ".join(vdf_columns_names(by, self)) + " ", by if polynomial_order <= 0: if two_sided: if period == 1: window = (-1, 1) else: if period % 2 == 0: window = (-period / 2 + 1, period / 2) else: window = (int(-period / 2), int(period / 2)) else: if period == 1: window = (-2, 0) else: window = (-period + 1, 0) vdf_tmp.rolling("avg", window, column, by_tmp, ts, trend_name) else: vdf_poly = vdf_tmp.copy() X = [] for i in range(1, polynomial_order + 1): vdf_poly[f"t_{i}"] = f"POWER(ROW_NUMBER() OVER ({by}ORDER BY {ts}), {i})" X += [f"t_{i}"] schema = vdf_poly._VERTICAPY_VARIABLES_["schema_writing"] if not (schema): schema = vdf_poly._VERTICAPY_VARIABLES_["schema"] if not (schema): schema = "public" from verticapy.learn.linear_model import LinearRegression model = LinearRegression(name="{}.VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format(schema, get_session(vdf_poly._VERTICAPY_VARIABLES_["cursor"])), cursor=vdf_poly._VERTICAPY_VARIABLES_["cursor"], solver="bfgs", max_iter=100, tol=1e-6,) model.drop() model.fit(vdf_poly, X, column) coefficients = model.coef_["coefficient"] coefficients = [str(coefficients[0])] + [f"{coefficients[i]} * POWER(ROW_NUMBER() OVER({by}ORDER BY {ts}), {i})" if i != 1 else f"{coefficients[1]} * ROW_NUMBER() OVER({by}ORDER BY {ts})" for i in range(1, polynomial_order + 1)] vdf_tmp[trend_name] = " + ".join(coefficients) model.drop() if mult: vdf_tmp[seasonal_name] = f'{column} / NULLIFZERO("{trend_name}")' else: vdf_tmp[seasonal_name] = vdf_tmp[column] - vdf_tmp[trend_name] if period <= 0: acf = vdf_tmp.acf(column=seasonal_name, ts=ts, p=23, acf_type="heatmap", show=False) period = int(acf["index"][1].split("_")[1]) if period == 1: period = int(acf["index"][2].split("_")[1]) vdf_tmp["row_number_id"] = f"MOD(ROW_NUMBER() OVER ({by} ORDER BY {ts}), {period})" if mult: vdf_tmp[ seasonal_name ] = f"AVG({seasonal_name}) OVER (PARTITION BY row_number_id) / NULLIFZERO(AVG({seasonal_name}) OVER ())" else: vdf_tmp[ seasonal_name ] = f"AVG({seasonal_name}) OVER (PARTITION BY row_number_id) - AVG({seasonal_name}) OVER ()" if estimate_seasonality: vdf_seasonality = vdf_tmp.copy() vdf_seasonality["t_cos"] = f"COS(2 * PI() * ROW_NUMBER() OVER ({by}ORDER BY {ts}) / {period})" vdf_seasonality["t_sin"] = f"SIN(2 * PI() * ROW_NUMBER() OVER ({by}ORDER BY {ts}) / {period})" X = ["t_cos", "t_sin",] schema = vdf_seasonality._VERTICAPY_VARIABLES_["schema_writing"] if not (schema): schema = vdf_seasonality._VERTICAPY_VARIABLES_["schema"] if not (schema): schema = "public" from verticapy.learn.linear_model import LinearRegression model = LinearRegression(name="{}.VERTICAPY_TEMP_MODEL_LINEAR_REGRESSION_{}".format(schema, get_session(vdf_seasonality._VERTICAPY_VARIABLES_["cursor"])), cursor=vdf_seasonality._VERTICAPY_VARIABLES_["cursor"], solver="bfgs", max_iter=100, tol=1e-6,) model.drop() model.fit(vdf_seasonality, X, seasonal_name) coefficients = model.coef_["coefficient"] vdf_tmp[seasonal_name] = f"{coefficients[0]} + {coefficients[1]} * COS(2 * PI() * ROW_NUMBER() OVER ({by}ORDER BY {ts}) / {period}) + {coefficients[2]} * SIN(2 * PI() * ROW_NUMBER() OVER ({by}ORDER BY {ts}) / {period})" model.drop() if mult: vdf_tmp[ epsilon_name ] = f'{column} / NULLIFZERO("{trend_name}") / NULLIFZERO("{seasonal_name}")' else: vdf_tmp[epsilon_name] = ( vdf_tmp[column] - vdf_tmp[trend_name] - vdf_tmp[seasonal_name] ) vdf_tmp["row_number_id"].drop() return vdf_tmp
def test_repr(self, model): assert "|coefficient|std_err |t_value |p_value" in model.__repr__() model_repr = LinearRegression("lin_repr") model_repr.drop() assert model_repr.__repr__() == "<LinearRegression>"