def main(queryfile: str, features: Optional[str] = None, halving: Optional[bool] = False, save: Optional[bool] = True): service = GridSearchService() models = ModelService() with open(queryfile, 'r') as f: query = json.load(f) search_models = models.query_models(query) logging.info("[i] {} models to train".format(len(search_models))) for i, m in enumerate(search_models): # if m.parameters: # logging.info("==[{}/{}]== MODEL: {} {} {} {} ==> SKIP".format(i+1, len(search_models), m.symbol, m.dataset, m.target, m.pipeline)) # continue # Skip this as search has already been performed logging.info("==[{}/{}]== MODEL: {} {} {} {} =====".format( i + 1, len(search_models), m.symbol, m.dataset, m.target, m.pipeline)) mp = service.create_parameters_search(m, split=0.7, features=features) logging.info("[{}] Start random search".format(get_timestamp())) mp = service.random_search(m, mp, sync=True, verbose=1, n_jobs=8, halving=halving, save=save) logging.info("[{}] End random search".format(get_timestamp()))
def main(pipeline: str, dataset: str, symbol: str, target: str, window: int, features: Optional[str] = None, parameters: Optional[str] = None, save: Optional[bool] = True): models = ModelService() try: logging.info( f"[{get_timestamp()}] {pipeline}({dataset}.{symbol}, W={window}) -> {target}," f"Features: {features}, Parameters: {parameters}, {'Persisted' if save else 'Non-Persisted'}" ) models.test_model_new(pipeline=pipeline, dataset=dataset, symbol=symbol, target=target, features=features, parameters=parameters, split=0.7, window={'days': window}, save=save) except MessageException as e: logging.error("[!] TEST FAILED, MESSAGE: " + e.message) pass except Exception as e: logging.exception("[!] TEST FAILED: " + str(e)) pass else: logging.info( f"[{get_timestamp()}] {pipeline}({dataset}.{symbol}, W={window}) -> {target} DONE" )
def main(): models = ModelService() datasets = DatasetService() query = { "dataset": "merged_new", "target": "class" } all_models = models.query_models(query=query) for m in all_models: ds = datasets.get_dataset(name=m.dataset, symbol=m.symbol) fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target=m.target) if not fs: logging.error(f"Dataset {m.dataset}{m.symbol} -> {m.target} does not have feature selection") continue if not m.parameters: logging.error(f"Model {m.pipeline}({m.dataset}{m.symbol}) -> {m.target} does not have parameters") continue for mp in m.parameters: count = 0 for f in mp.features: if not f in fs.features: logging.error(f"Model {m.pipeline}({m.dataset}{m.symbol}) -> {m.target} parameter search done without fixing features!") else: count += 1 logging.info(f"Model {m.pipeline}({m.dataset}{m.symbol}) -> {m.target} GRIDSEARCH {mp.parameter_search_method} done with {count} features")
def main(queryfile: str): with open(queryfile, 'r') as f: query = json.load(f) service = ModelService() print(f"Clearing {query}") time.sleep(5) service.clear_tests(query)
def main(dataset: str, pipeline: str): models = ModelService() query = {'type': "FEATURES"} if dataset != 'all': query['name'] = dataset items = models.create_classification_models(query=query, pipeline=pipeline) print(items)
def model_index(features: Optional[bool] = True, parameters: Optional[bool] = True, tests: Optional[bool] = True, query: Optional[dict] = Body(...), service: ModelService = Depends(ModelService)): result = {} if features: result['features'] = service.clear_features(query) if parameters: result['parameters'] = service.clear_parameters(query) if tests: result['tests'] = service.clear_tests(query) return result
def feature_selection( model_id: str, method: str, split: Optional[float] = 0.7, batch: Optional[str] = None, task_key: Optional[str] = None, sync: Optional[bool] = False, model_service: ModelService = Depends(ModelService), service: FeatureSelectionService = Depends(FeatureSelectionService), tasks: TaskService = Depends(TaskService)): try: model = model_service.get_model(model_id) mf = service.create_features_search(model, split, method, task_key=task_key) if sync: return service.feature_selection(model, mf, sync=True) return tasks.send(task_name='featureselection', task_args={ 'model': model.dict(), 'search_parameters': mf.dict() }, name='feature_selection-{}-{}-{}-{}'.format( model.symbol, model.pipeline, model.dataset, model.target), batch=batch) except MessageException as e: raise HTTPException(status_code=400, detail=e.message)
def grid_search_batch(batch: Optional[str] = None, task_key: Optional[str] = None, split: Optional[float] = 0.7, query: dict = Body(...), model_service: ModelService = Depends(ModelService), service: GridSearchService = Depends(GridSearchService), tasks: TaskService = Depends(TaskService)): try: models = model_service.query_models(query) tests = [(model, service.create_parameters_search(model, split, task_key=task_key)) for model in models] return [ tasks.send(task_name='gridsearch', task_args={ 'model': model.dict(), 'search_parameters': search_parameters.dict() }, name='grid_search-{}-{}-{}-{}'.format( model.symbol, model.pipeline, model.dataset, model.target), batch=batch, countdown=30) for model, search_parameters in tests ] except MessageException as e: raise HTTPException(status_code=400, detail=e.message)
def grid_search(model_id: str, split: Optional[float] = 0.7, batch: Optional[str] = None, task_key: Optional[str] = None, sync: Optional[bool] = False, model_service: ModelService = Depends(ModelService), service: GridSearchService = Depends(GridSearchService), tasks: TaskService = Depends(TaskService)): try: model = model_service.get_model(model_id) parameters = service.create_parameters_search(model, split, task_key=task_key) if sync: return service.grid_search(model, parameters, sync=True) return tasks.send(task_name='gridsearch', task_args={ 'model': model.dict(), 'search_parameters': parameters.dict() }, name='grid_search-{}-{}-{}-{}'.format( model.symbol, model.pipeline, model.dataset, model.target), batch=batch) except MessageException as e: raise HTTPException(status_code=400, detail=e.message)
def main(dataset: str, target: str, pipeline: str): shapes = [] ds_service = DatasetService() m_service = ModelService() for symbol in SYMBOLS: print(f"Exporting shap dataframes for symbol {symbol}") ds = ds_service.get_dataset(name=dataset, symbol=symbol) fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target=target) X_all = ds_service.get_dataset_features(ds=ds, columns=fs.features) y_all = ds_service.get_dataset_target(ds=ds, name=target) model = m_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) for t in model.tests: print(f"Loading estimators for test {t.window}") estimators = ModelService.load_test_estimators(model=model, mt=t) shaps = [] print(f"Calculating shap values...") for est in tqdm(estimators): est_class = y_all.loc[est.day] shap_v, shap_exp = get_shap_values(estimator=est, X=X_all.loc[est.day], X_train=est.train_x, bytes=False) df = pd.DataFrame([shap_v], index=[pd.to_datetime(est.day)], columns=X_all.columns) df['label'] = y_all.loc[est.day] df['shap_expected'] = shap_exp shaps.append(df) print("Exporting dataframe..") cdf = pd.concat(shaps, axis='index') os.makedirs(f"data/shap_values/{dataset}/{target}/{pipeline}/", exist_ok=True) cdf.to_csv( f"data/shap_values/{dataset}/{target}/{pipeline}/shap_test_{symbol}_Wdays{t.window['days']}.csv", index_label='time') print("Exported.") # # Load day estimator # est = load_estimator() print(f"Plotted {symbol}")
def test_model(model_id: str, sync: Optional[bool] = False, test: ModelTest = Body(...), tasks: TaskService = Depends(TaskService), service: ModelService = Depends(ModelService)): try: model = service.get_model(model_id) if sync: return service.test_model(model, test) return tasks.send(task_name='testmodel', task_args={ 'model': model.dict(), 'test': test.dict() }, name='model_test-{}-{}-{}-{}'.format( model.symbol, model.pipeline, model.dataset, model.target)) except MessageException as e: raise HTTPException(status_code=400, detail=e.message)
def main(dataset: str, target: str): service = FeatureSelectionService() models = ModelService() datasets = DatasetService() query = {"dataset": dataset, "target": target} # Clear feature search results from models models.clear_features(query) #search_models = models.query_models(query) # logging.info("[i] {} models for feature selection".format(len(search_models))) # for i, m in enumerate(search_models): symbols = datasets.get_dataset_symbols(dataset) for i, sym in enumerate(symbols): logging.info("==[{}/{}]== Dataset: {} {} {} =====".format( i + 1, len(symbols), sym, dataset, target)) mf = service.create_features_search(target=target, dataset=dataset, symbol=sym, split=0.7, method='importances') logging.info("[{}] Start feature search".format(get_timestamp())) mf = service.feature_selection(mf, sync=True) logging.info("[{}] End feature search".format(get_timestamp()))
def feature_selection_batch( method: str, batch: Optional[str] = None, task_key: Optional[str] = None, split: Optional[float] = 0.7, query: dict = Body(...), model_service: ModelService = Depends(ModelService), service: FeatureSelectionService = Depends(FeatureSelectionService), tasks: TaskService = Depends(TaskService)): try: models = model_service.query_models(query) # This will only keep 1 copy for each (symbol, dataset, target) tuple d_models = { '{}-{}-{}'.format(m.symbol, m.dataset, m.target): m for m in models } models = [v for k, v in d_models.items()] def get_name_from_model(_model): return 'feature_selection-{}-{}-{}-{}'.format( _model.symbol, _model.pipeline, _model.dataset, _model.target) tests = [(model, service.create_features_search(model, split, method, task_key=task_key)) for model in models] return [ tasks.send(task_name='featureselection', task_args={ 'model': model.dict(), 'search_parameters': search_parameters.dict() }, name=get_name_from_model(model), batch=batch, countdown=30) for i, (model, search_parameters) in enumerate(tests) ] except MessageException as e: raise HTTPException(status_code=400, detail=e.message)
def main(dataset: str, target: str): num_shap_plots = 3 shap_show_count = 10 ds_service = DatasetService() m_service = ModelService() for pipeline in PIPELINES: for symbol in SYMBOLS: print( f"Plotting shap dataframes for pipeline {pipeline} symbol {symbol}" ) ds = ds_service.get_dataset(name=dataset, symbol=symbol) fs = DatasetService.get_feature_selection( ds=ds, method='importances_shap', target=target) X_all = ds_service.get_dataset_features(ds=ds, columns=fs.features) y_all = ds_service.get_dataset_target(ds=ds, name=target) model = m_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) for t in model.tests: placeholder = "{label}" csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/shap_training_window_{symbol}_{placeholder}_Wdays{t.window['days']}_.csv" expected_csv_name = csv_name.format(label='SHAP_expected') print(f"Loading results for test {t.window}") results = ModelService.parse_test_results(test=t) exp_shap_df = pd.read_csv(expected_csv_name, index_col='time', parse_dates=True) for cls, label in enumerate(["SELL", "HOLD", "BUY"]): class_csv_name = csv_name.format(label=label) cls_shap_df = pd.read_csv(class_csv_name, index_col='time', parse_dates=True) cls_shap_df = cls_shap_df.loc[t.test_interval.begin:t. test_interval.end] x_train = X_all.loc[cls_shap_df.index] chunk_size = int(cls_shap_df.shape[0] / num_shap_plots) fig = plt.figure(constrained_layout=True, figsize=(100, 50), dpi=300) # gs = GridSpec(3, num_shap_plots, figure=fig, wspace=1.5, hspace=0.3) precision_ax = fig.add_subplot(gs[0, :]) shap_values_ax = fig.add_subplot(gs[1, :]) beeswarms_axs = [ fig.add_subplot(gs[2, i]) for i in range(num_shap_plots) ] #format_axes(fig) shap_plot_labels = set() first_shap_day = results.iloc[0]['time'].replace( '+00:00', '').replace('T', '').replace(':', '').replace('-', '') middle_shap_day = results.iloc[int( results.shape[0] / 2)]['time'].replace( '+00:00', '').replace('T', '').replace(':', '').replace('-', '') last_shap_day = results.iloc[-1]['time'].replace( '+00:00', '').replace('T', '').replace(':', '').replace('-', '') for idx, dayname in enumerate( [first_shap_day, middle_shap_day, last_shap_day]): day_csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/daily/shap_training_window_{symbol}_{label}_Wdays{t.window['days']}_DAY{dayname}.csv" # Plot each section's SHAP values cdf_subset = pd.read_csv(day_csv_name, index_col='time', parse_dates=True) train_subset = X_all.loc[cdf_subset.index] # Get a rank of feature labels based on this section's shap values abs_mean_shap = cdf_subset.abs().mean(axis='index') abs_mean_rank = abs_mean_shap.sort_values( ascending=False)[:shap_show_count] for l in abs_mean_rank.index: # Save labels for features in the top-N shap_plot_labels.add(l) # Plot this section's SHAP values plt.sca(beeswarms_axs[idx]) shap.summary_plot(cdf_subset.values, train_subset, max_display=shap_show_count, show=False, color_bar=False, sort=True) min_date = cdf_subset.index.min().to_pydatetime() max_date = cdf_subset.index.max().to_pydatetime( ) + timedelta(days=1) min_date_f = min_date.strftime("%Y/%m/%d") max_date_f = max_date.strftime("%Y/%m/%d") beeswarms_axs[idx].set_xlabel( f"SHAP values\nWindow: {min_date_f} - {max_date_f}", fontsize=8) beeswarms_axs[idx].tick_params(axis='y', which='major', labelsize=6) beeswarms_axs[idx].tick_params(axis='x', which='major', labelsize=8) # Plot shap values day_csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/shap_training_window_{symbol}_{label}_Wdays{t.window['days']}_.csv" plot_cls_shap_df = pd.read_csv(day_csv_name, index_col='time', parse_dates=True) def get_spread(series): return np.abs(series.max() - series.min()) plot_rank = plot_cls_shap_df[list(shap_plot_labels)].apply( get_spread, axis='index').sort_values( ascending=False)[:shap_show_count] plot_cls_shap_df['xlabel'] = [ t.to_pydatetime().strftime("%Y/%m/%d") for t in plot_cls_shap_df.index ] shap_ax = plot_cls_shap_df.plot( x='xlabel', y=[c for c in plot_rank.index], kind='line', ax=shap_values_ax, legend=False, xlabel='') patches, labels = shap_ax.get_legend_handles_labels() shap_ax.legend(patches, labels, loc='center left', bbox_to_anchor=(1, 0.5), prop={'size': 6}) shap_ax.tick_params(axis='x', which='major', labelsize=8) shap_ax.set_ylabel('mean(|SHAP|)', fontsize=6) #shap_ax.tick_params(labelbottom=False, labelleft=False) # Get Metrics scores dataframe cri_df = get_metrics_df(results).rolling( 7, min_periods=1).mean() cri_df['xlabel'] = [ t.to_pydatetime().strftime("%Y/%m/%d") for t in cri_df.index ] cri_ax = cri_df.plot(x='xlabel', y=f"pre_{cls}", kind='line', ax=precision_ax, legend=False, xlabel='') patches, labels = cri_ax.get_legend_handles_labels() cri_ax.legend(patches, labels, loc='center left', bbox_to_anchor=(1, 0.5), prop={'size': 6}) cri_ax.set_ylabel('mean(precision)', fontsize=6) cri_ax.tick_params(labelbottom=False, labelleft=True) min_date = cri_df.index.min().to_pydatetime().strftime( "%Y/%m/%d") max_date = cri_df.index.max().to_pydatetime().strftime( "%Y/%m/%d") window = t.window['days'] fig.suptitle( f"{symbol}, {pipeline}, W={window}D, Class {label}, From {min_date} to {max_date}" ) # fig.show() os.makedirs(f"images/shap-test-final/", exist_ok=True) plt.savefig( f"images/shap-test-final/{pipeline}_W{window}D_{dataset}_{target}_{symbol}_{label}.png", dpi='figure') plt.close() print(f"{label} OK") print(f"Exported symbol {symbol}.") # # Load day estimator # est = load_estimator() print(f"Plotted {symbol}")
def main(dataset: str): ds = DatasetService() ms = ModelService() ts = TradingService() logs = [] for pipeline in PIPELINES: for symbol in SYMBOLS: for window in WINDOWS: print( f"PIPELINE: {pipeline} SYMBOL: {symbol} WINDOW: {window}") ohlcv_ds = ds.get_dataset('ohlcv', symbol=symbol) test = ms.get_test(pipeline=pipeline, dataset=dataset, target='class', symbol=symbol, window=window) if not test: print( f"Test {pipeline}.{dataset}.class for {symbol} on window {window} not found!" ) logs.append( f"MISSING_TEST {pipeline} {dataset} {symbol} class {window} --features importances_shap --parameters gridsearch\n" ) continue asset = ts.get_asset(pipeline=pipeline, dataset=dataset, target='class', symbol=symbol, window=window, create=False) if not asset: print( f"Asset {pipeline}.{dataset}.class for {symbol} on window {window} not found!" ) logs.append( f"MISSING_ASSET {pipeline} {dataset} {symbol} {window}\n" ) continue equity = TradingService.parse_equity_df(asset=asset) buy_and_hold = TradingService.parse_baseline_df( asset=asset, name='buy_and_hold') orders = TradingService.parse_orders_df(asset=asset) # Map order position_id to numbers so we don't get a mess in the graph position_uids = set(orders.position_id.values) for i, uid in enumerate(position_uids): orders.position_id.replace(to_replace=uid, value=i, inplace=True) ohlcv = ds.get_dataset_features(ohlcv_ds, begin=test.test_interval.begin, end=test.test_interval.end) test_results = ModelService.parse_test_results(test).iloc[:-1] # Mask predictions with low value minus a certain amount signals_level_diff = ohlcv.low * 10 / 100 signals_level = ohlcv.low - signals_level_diff enc_pred = onehot_target( test_results.predicted, labels=["is_sell", "is_hold", "is_buy"], fill=False) # In case of classifier bias (due to input bias) some classes are ignored. # In such cases, enc_pred won't contain the ignored classes. # Add them back by nan-filling (never selected) if hasattr(enc_pred, 'is_sell'): use_idx = enc_pred.is_sell > 0 enc_pred.is_sell.mask( use_idx, other=signals_level.loc[enc_pred.index], inplace=True) else: enc_pred['is_sell'] = pd.Series(np.nan, index=enc_pred.index) if hasattr(enc_pred, 'is_hold'): enc_pred.is_hold.mask( enc_pred.is_hold > 0, other=signals_level.loc[enc_pred.index], inplace=True) else: enc_pred['is_hold'] = pd.Series(np.nan, index=enc_pred.index) if hasattr(enc_pred, 'is_buy'): enc_pred.is_buy.mask( enc_pred.is_buy > 0, other=signals_level.loc[enc_pred.index], inplace=True) else: enc_pred['is_buy'] = pd.Series(np.nan, index=enc_pred.index) # Get unique years in index to split plots in smaller scale unique_years = ohlcv.index.year.unique() for year in unique_years: year_ohlcv = ohlcv[ohlcv.index.year == year] year_pred = enc_pred[enc_pred.index.year == year] year_equity = equity[equity.index.year == year] year_buy_and_hodl = buy_and_hold[buy_and_hold.index.year == year] year_orders = orders[orders.index.year == year] unique_quarters = year_ohlcv.index.quarter.unique() for quarter in unique_quarters: q_ohlcv = year_ohlcv[year_ohlcv.index.quarter == quarter] q_pred = year_pred[year_pred.index.quarter == quarter] q_equity = year_equity[year_equity.index.quarter == quarter] q_orders = year_orders[year_orders.index.quarter == quarter] q_buy_and_hodl = year_buy_and_hodl[ year_buy_and_hodl.index.quarter == quarter] #f"{ohlcv_ds.symbol}, {year} - Q{quarter}, 1D", 'Trades', 'Equity' img_path = f"images/backtests-final/{pipeline}-{dataset}-class-W{window}/{symbol}/" img_name = f"trades-{year}-Q{quarter}.png" if os.path.exists(f"{img_path}/{img_name}"): print(f"[SKIP] File exists {img_path}/{img_name}") continue make_plot( ohlcv=q_ohlcv, orders=q_orders, equity=q_equity, baselines=[('Buy and Hold', q_buy_and_hodl)], pred=q_pred, signals_title= f"{ohlcv_ds.symbol}, {pipeline}, W={window}D, {year} - Q{quarter}, 1D", img_path=img_path, img_name=img_name, bollinger=True) print( f"{year}-Q{quarter} saved to {img_path}{img_name}") with open(f"trading_plotly.{dataset}.log", "w") as f: f.writelines(logs) print("Logs saved")
def main(dataset: str, target: str, pipeline: str): shapes = [] ds_service = DatasetService() m_service = ModelService() for symbol in SYMBOLS: print(f"Exporting shap dataframes for symbol {symbol}") ds = ds_service.get_dataset(name=dataset, symbol=symbol) fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target=target) X_all = ds_service.get_dataset_features(ds=ds, columns=fs.features) y_all = ds_service.get_dataset_target(ds=ds, name=target) model = m_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) for t in model.tests: os.makedirs( f"data/shap_values/{dataset}/{target}/{pipeline}/daily", exist_ok=True) placeholder = "{label}" csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/shap_training_window_{symbol}_{placeholder}_Wdays{t.window['days']}_.csv" day_csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/daily/shap_training_window_{symbol}_{placeholder}_Wdays{t.window['days']}_" print(f"Loading estimators for test {t.window}") estimators = ModelService.load_test_estimators(model=model, mt=t) results = ModelService.parse_test_results(test=t) shaps = [[], [], []] X_test = X_all.loc[t.test_interval.begin:t.test_interval.end] shap_expected = [] print(f"Calculating shap values") shap_abs_mean = [pd.DataFrame(), pd.DataFrame(), pd.DataFrame()] for est in tqdm(estimators): est_class = y_all.loc[est.day] training_data = est.train_x.astype(np.float64).fillna(value=0) shap_v, shap_exp = get_shap_values(estimator=est.named_steps.c, X=training_data, X_train=training_data, bytes=False) if isinstance(shap_exp, float): shap_expected.append([est.day] + [0, 0, shap_exp]) else: shap_expected.append([est.day] + [v for v in shap_exp]) for cls, label in enumerate(["SELL", "HOLD", "BUY"]): df = pd.DataFrame(shap_v[cls], index=est.train_x.index, columns=est.train_x.columns) # if not shaps[cls]: # If list is empty, append whole df # shaps[cls].append(df) # else: # shaps[cls].append(df.iloc[-1:]) # otherwise only append new row (sliding window) # Save shap values dataframe for each day dayname = est.day.replace('+00:00', '').replace('T', '').replace( ':', '').replace('-', '') day_class_csv_name = day_csv_name.format( label=label) + f"DAY{dayname}.csv" df.to_csv(day_class_csv_name, index_label='time') # Process data for next plot df_abs_mean = df.abs().mean().to_dict() df_abs_mean['time'] = est.day shaps[cls].append(df_abs_mean) # print(shap_abs_mean.head()) # Merge shap values in an unique dataframe and save to csv for each class for cls, label in enumerate(["SELL", "HOLD", "BUY"]): class_csv_name = csv_name.format(label=label) print( f"Exporting dataframe for class {label} -> {class_csv_name}" ) # cdf = pd.concat(shaps[cls], axis='index') cdf = pd.DataFrame.from_records(shaps[cls]) cdf.index = pd.to_datetime(cdf.time) cdf = cdf[cdf.columns.difference(['time'])] cdf.to_csv(class_csv_name, index_label='time') expected_csv_name = csv_name.format(label='SHAP_expected') print( f"Exporting expected values dataframe -> {expected_csv_name}") edf = pd.DataFrame( shap_expected, columns=[ "time", "shap_expected_sell", "shap_expected_hold", "shap_expected_buy" ], ) edf.to_csv(expected_csv_name, index_label='time') print(f"Exported symbol {symbol}.") # # Load day estimator # est = load_estimator() print(f"Plotted {symbol}")
def task_test_model(req: dict): model = Model(**req['model']) test = ModelTest(**req['test']) model_service = ModelService() mt = model_service.test_model(model, test) return mt.dict()
def main(pipeline: str, dataset: str, symbol: str, window: int): ds = DatasetService() ms = ModelService() ts = TradingService() ohlcv_ds = ds.get_dataset('ohlcv', symbol=symbol) asset = ts.get_asset(pipeline=pipeline, dataset=dataset, target='class', symbol=symbol, window=window, create=False) if not asset: print( f"Asset {pipeline}.{dataset}.class for {symbol} on window {window} not found!" ) return test = ms.get_test(pipeline=pipeline, dataset=dataset, target='class', symbol=symbol, window=window) if not test: print( f"Test {pipeline}.{dataset}.class for {symbol} on window {window} not found!" ) # ohlcv = ohlcv.loc[test.test_interval.begin:test.test_interval.end] ohlcv = ds.get_dataset_features(ohlcv_ds, begin=test.test_interval.begin, end=test.test_interval.end) test_results = ModelService.parse_test_results(test).iloc[:-1] enc_label = onehot_target(test_results.label, labels=["is_sell", "is_hold", "is_buy"], fill=False) enc_pred = onehot_target(test_results.predicted, labels=["is_sell", "is_hold", "is_buy"], fill=False) # Mask predictions with low value minus a certain amount signals_level_diff = ohlcv.low * 10 / 100 signals_level = ohlcv.low - signals_level_diff #signals_level = ohlcv.low enc_pred.is_sell.mask(enc_pred.is_sell > 0, other=signals_level, inplace=True) enc_pred.is_hold.mask(enc_pred.is_hold > 0, other=signals_level, inplace=True) enc_pred.is_buy.mask(enc_pred.is_buy > 0, other=signals_level, inplace=True) # Get unique years in index to split plots in smaller scale unique_years = ohlcv.index.year.unique() for year in unique_years: year_pred = enc_pred[enc_pred.index.year == year] year_ohlcv = ohlcv[ohlcv.index.year == year] # Set up xticks daysToIndex = { ts.to_pydatetime(): i for i, ts in enumerate(year_ohlcv.index) } days = [i for i in daysToIndex.values()] labels = [ ts.to_pydatetime().strftime("%Y-%m-%d") for ts in year_ohlcv.index ] # Setup matplotfinance styles and figure s = mpf.make_mpf_style( base_mpf_style='binance') # , rc={'font.size': 6} fig = mpf.figure( figsize=(16, 8), style=s) # pass in the self defined style to the whole canvas fig.suptitle(f"{ohlcv_ds.symbol}, {year}, 1D") ax = fig.add_subplot(3, 1, (1, 2)) # main candle stick chart subplot av = fig.add_subplot(3, 1, 3, sharex=ax) # volume candles subplot # Setup horizontal grids ax.grid(axis='x', color='0.5', linestyle='--') av.grid(axis='x', color='0.5', linestyle='--') # for a in [ax, av]: # a.set_xticks(ticks=days) # a.set_xticklabels(labels=labels) # a.tick_params(axis='x', labelrotation=90) apds = [ # mpf.make_addplot(tcdf) # Predictions mpf.make_addplot(year_ohlcv.close, ax=ax, type='line', color=(0.5, 0.5, 0.5, 0.05)), mpf.make_addplot(year_pred.is_sell, ax=ax, type='scatter', marker='v', color='red'), mpf.make_addplot(year_pred.is_hold, ax=ax, type='scatter', marker='_', color='silver'), mpf.make_addplot(year_pred.is_buy, ax=ax, type='scatter', marker='^', color='lime'), ] mpf.plot( year_ohlcv, type='candle', style=s, #ylabel='Price ($)', ax=ax, volume=av, #ylabel_lower='Volume', show_nontrading=True, addplot=apds, returnfig=True) fig.autofmt_xdate() fig.tight_layout() plt.show() print("Done")
def __init__(self): self.model_repo = ModelRepository() self.model_service = ModelService() self.dataset_service = DatasetService()
def model_index(service: ModelService = Depends(ModelService)): return service.all()
def main(dataset: str, target: str, pipeline: str): hierarchy = load_hierarchy(f"{dataset}_{target}_feature_hierarchy.yml") hdf = pd.DataFrame(hierarchy) num_shap_plots = 3 shap_show_count = 10 ds_service = DatasetService() m_service = ModelService() for symbol in SYMBOLS: print(f"Plotting shap dataframes for symbol {symbol}") ds = ds_service.get_dataset(name=dataset, symbol=symbol) fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target=target) X_all = ds_service.get_dataset_features(ds=ds, columns=fs.features) y_all = ds_service.get_dataset_target(ds=ds, name=target) model = m_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) for t in model.tests: os.makedirs( f"images/shap-test-hierarchy/{dataset}/{target}/{pipeline}/", exist_ok=True) placeholder = "{label}" csv_name = f"data/shap_values/{dataset}/{target}/{pipeline}/shap_training_window_{symbol}_{placeholder}_Wdays{t.window['days']}_.csv" expected_csv_name = csv_name.format(label='SHAP_expected') print(f"Loading results for test {t.window}") results = ModelService.parse_test_results(test=t) exp_shap_df = pd.read_csv(expected_csv_name, index_col='time', parse_dates=True) for cls, label in enumerate(["SELL", "HOLD", "BUY"]): class_csv_name = csv_name.format(label=label) cls_shap_df = pd.read_csv(class_csv_name, index_col='time', parse_dates=True) cls_shap_df = cls_shap_df.loc[t.test_interval.begin:t. test_interval.end] x_train = X_all.loc[cls_shap_df.index] chunk_size = int(cls_shap_df.shape[0] / num_shap_plots) # fig = plt.figure(constrained_layout=True, figsize=(100, 50), dpi=300) # # gs = GridSpec(3, num_shap_plots, figure=fig, wspace=1.5, hspace=0.3) # precision_ax = fig.add_subplot(gs[0, :]) # shap_values_ax = fig.add_subplot(gs[1, :]) # beeswarms_axs = [fig.add_subplot(gs[2, i]) for i in range(num_shap_plots)] # #format_axes(fig) # shap_plot_labels = set() # for idx, start in enumerate(range(0, cls_shap_df.shape[0], chunk_size)): # end = start + chunk_size # left = cls_shap_df.shape[0] - end # if left > 0 and left < chunk_size: # end += left # elif left < 0: # break # # Plot each section's SHAP values # cdf_subset = cls_shap_df.iloc[start:end] # train_subset = x_train.iloc[start:end] # # # Get a rank of feature labels based on this section's shap values # abs_mean_shap = cdf_subset.abs().mean(axis='index') # abs_mean_rank = abs_mean_shap.sort_values(ascending=False)[:shap_show_count] # for l in abs_mean_rank.index: # # Save labels for features in the top-N # shap_plot_labels.add(l) # # # Plot this section's SHAP values # plt.sca(beeswarms_axs[idx]) # shap.summary_plot( # cdf_subset.values, # train_subset, # max_display=shap_show_count, # show=False, # color_bar=False, # sort=True # ) # min_date = cdf_subset.index.min().to_pydatetime().strftime("%Y/%m/%d") # max_date = cdf_subset.index.max().to_pydatetime().strftime("%Y/%m/%d") # beeswarms_axs[idx].set_xlabel(f"SHAP values\n{min_date} - {max_date}", fontsize=8) # beeswarms_axs[idx].tick_params(axis='y', which='major', labelsize=6) # beeswarms_axs[idx].tick_params(axis='x', which='major', labelsize=8) # # Plot shap values # plot_cls_shap_df = cls_shap_df.abs().rolling(7, min_periods=1).mean() # def get_spread(series): # return np.abs(series.max() - series.min()) # plot_rank = plot_cls_shap_df[list(shap_plot_labels)].apply(get_spread, axis='index').sort_values(ascending=False)[:shap_show_count] # plot_cls_shap_df['xlabel'] = [t.to_pydatetime().strftime("%Y/%m/%d") for t in plot_cls_shap_df.index] # shap_ax = plot_cls_shap_df.plot( # x='xlabel', # y=[c for c in plot_rank.index], # kind='line', # ax=shap_values_ax, # legend=False, # xlabel='' # ) # patches, labels = shap_ax.get_legend_handles_labels() # shap_ax.legend( # patches, labels, # loc='center left', bbox_to_anchor=(1, 0.5), prop={'size': 6} # ) # shap_ax.tick_params(axis='x', which='major', labelsize=8) # shap_ax.set_ylabel('mean(|SHAP|)', fontsize=6) # #shap_ax.tick_params(labelbottom=False, labelleft=False) # # # Get Metrics scores dataframe # cri_df = get_metrics_df(results).rolling(7, min_periods=1).mean() # cri_df['xlabel'] = [t.to_pydatetime().strftime("%Y/%m/%d") for t in cri_df.index] # cri_ax = cri_df.plot( # x='xlabel', # y=f"pre_{cls}", # kind='line', # ax=precision_ax, # legend=False, # xlabel='' # ) # patches, labels = cri_ax.get_legend_handles_labels() # cri_ax.legend( # patches, labels, # loc='center left', bbox_to_anchor=(1, 0.5), prop={'size': 6} # ) # cri_ax.set_ylabel('mean(precision)', fontsize=6) # cri_ax.tick_params(labelbottom=False, labelleft=True) # # min_date = cri_df.index.min().to_pydatetime().strftime("%Y/%m/%d") # max_date = cri_df.index.max().to_pydatetime().strftime("%Y/%m/%d") # fig.suptitle(f"{pipeline}, {symbol}, class {label} tests from {min_date} to {max_date}") # # # fig.show() # plt.savefig( # f"images/shap-test/{pipeline}_{dataset}_{target}_{symbol}_{label}.png", # dpi='figure' # ) # plt.close() print(f"{label} OK") print(f"Exported symbol {symbol}.") # # Load day estimator # est = load_estimator() print(f"Plotted {symbol}")
def main(queryfile: str, features: Optional[str] = None, parameters: Optional[str] = None, save: Optional[bool] = True): models = ModelService() with open(queryfile, 'r') as f: query = json.load(f) if save: models.clear_tests(query) test_models = models.query_models(query) logging.info("[i] {} models to test".format(len(test_models))) failed = [] for i, m in enumerate(test_models): logging.info("==[{}/{}]== MODEL: {} {} {} {} =====".format( i + 1, len(test_models), m.symbol, m.dataset, m.target, m.pipeline)) #t1 = models.create_model_test(model=m, split=0.7, step={'days': 1}, window={'days': 60}, parameters=parameters, features=features) t2 = models.create_model_test(model=m, split=0.7, step={'days': 1}, window={'days': 90}, parameters=parameters, features=features) t3 = models.create_model_test(model=m, split=0.7, step={'days': 1}, window={'days': 180}, parameters=parameters, features=features) t4 = models.create_model_test(model=m, split=0.7, step={'days': 1}, window={'days': 240}, parameters=parameters, features=features) try: # Test T1 # logging.info("[{}] {} Start T1".format(get_timestamp(), m.symbol)) # models.test_model(m, t1, sync=True) # Test T2 logging.info("[{}] {} Start T2".format(get_timestamp(), m.symbol)) models.test_model(m, t2, sync=True) # Test T3 logging.info("[{}] {} Start T3".format(get_timestamp(), m.symbol)) models.test_model(m, t3, sync=True) logging.info("[{}] {} Start T4".format(get_timestamp(), m.symbol)) models.test_model(m, t4, sync=True) except MessageException as e: logging.error("[!] " + e.message) # failed.append((m.dict(), t1.dict(), t2.dict(), t3.dict())) failed.append((m.dict(), t2.dict(), t3.dict(), t4.dict())) pass except Exception as e: logging.exception("[!] " + str(e)) # failed.append((m.dict(), t1.dict(), t2.dict(), t3.dict())) failed.append((m.dict(), t2.dict(), t3.dict(), t4.dict())) pass logging.info("[{}] Done".format(m.symbol)) with open('test-failed.json', 'w') as f: json.dump(failed, f)
def grid_search_new(self, symbol: str, dataset: str, target: str, pipeline: str, split: float, feature_selection_method: str, **kwargs): # Check if a model exists and has same search method existing_model = self.model_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) if existing_model: mp_exists = ModelService.get_model_parameters(existing_model, method='gridsearch') if mp_exists: if kwargs.get('replace'): self.model_service.remove_parameters(model=existing_model, method='gridsearch') else: if kwargs.get('save'): raise MessageException( f"Grid search already performed for {pipeline}({dataset}.{symbol}) -> {target}" ) # Retrieve dataset to use ds = self.dataset_service.get_dataset(dataset, symbol) # Determine cv_splits=K for K-fold cross validation based on dataset's sample count # Train-test split for each fold is 80% train, the lowest training window for accurate results is 30 samples # so we need X samples where X is given by the proportion: # 30/0.8 = X/1; X= 30/0.8 = 37.5 ~ 40 samples per fold X = 40 cv_splits = 5 # If samples per fold with 5-fold CV are too low, use 3-folds if ds.count / cv_splits < X: cv_splits = 3 # If samples are still too low, raise a value error if ds.count / cv_splits < X and not kwargs.get("permissive"): raise ValueError("Not enough samples to perform cross validation!") # Determine split indices based on dataset splits = DatasetService.get_train_test_split_indices(ds, split) cv_interval = splits['train'] # Load dataset features by applying a specified feature selection method X = self.dataset_service.get_dataset_features( ds=ds, begin=cv_interval['begin'], end=cv_interval['end'], method=feature_selection_method, target=target) y = self.dataset_service.get_target( name=target, symbol=symbol, begin=cv_interval['begin'], end=cv_interval['end'], ) # Check number of samples for each class in training data, if less than 3 instances are present for # each class, we're going to get a very unstable model (or no model at all for k-NN based algos) unique, counts = np.unique(y, return_counts=True) if len(unique) < 2: logging.error( "[{}-{}-{}-{}]Training data contains less than 2 classes: {}". format(symbol, dataset, target, pipeline, unique)) raise MessageException( "Training data contains less than 2 classes: {}".format( unique)) logging.info("Dataset loaded: X {} y {} (unique: {})".format( X.shape, y.shape, unique)) # Load pipeline algorithm and parameter grid pipeline_module = get_pipeline(pipeline) # Perform search gscv = GridSearchCV( estimator=pipeline_module.estimator, param_grid=kwargs.get('parameter_grid', pipeline_module.PARAMETER_GRID), # cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits), cv=StratifiedKFold(n_splits=cv_splits), scoring=get_precision_scorer(), verbose=kwargs.get("verbose", 0), n_jobs=kwargs.get("n_jobs", None), refit=False) mp = ModelParameters(cv_interval=splits['train'], cv_splits=cv_splits, task_key=kwargs.get('task_key', str(uuid4())), features=[c for c in X.columns], parameter_search_method='gridsearch') mp.start_at = get_timestamp() gscv.fit(X, y) mp.end_at = get_timestamp() # Collect results results_df = pd.DataFrame(gscv.cv_results_) mp.parameters = gscv.best_params_ mp.cv_results = results_df.loc[:, results_df.columns != 'params'].to_dict( 'records') tag = "{}-{}-{}-{}-{}".format(symbol, dataset, target, pipeline, dict_hash(mp.parameters)) mp.result_file = 'cv_results-{}.csv'.format(tag) # Is there an existing model for this search? model = Model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol, features=feature_selection_method) model.parameters.append(mp) self.model_repo.create(model) # Save grid search results on storage if kwargs.get('save', True): storage_service.upload_json_obj(mp.parameters, 'grid-search-results', 'parameters-{}.json'.format(tag)) storage_service.save_df(results_df, 'grid-search-results', mp.result_file) return mp
class GridSearchService: def __init__(self): self.model_repo = ModelRepository() self.model_service = ModelService() self.dataset_service = DatasetService() def create_parameters_search(self, model: Model, split: float, **kwargs) -> ModelParameters: ds = self.dataset_service.get_dataset(model.dataset, model.symbol) splits = DatasetService.get_train_test_split_indices(ds, split) # Features can either be a list of features to use, or a string # If it is a string, and it is "latest", pick the latest features = kwargs.get('features') # if isinstance(features, str) and features == 'latest': # if model.features: # features = model.features[-1].features # else: # features = None if features: target = kwargs.get('target', 'class') mf = DatasetService.get_feature_selection( ds=ds, method=kwargs.get('features'), target=target) if not mf: raise MessageException( f"Feature selection not found for {model.dataset}.{model.symbol} -> {target}!" ) features = mf.features # Determine K for K-fold cross validation based on dataset's sample count # Train-test split for each fold is 80% train, the lowest training window for accurate results is 30 samples # so we need X samples where X is given by the proportion: # 30/0.8 = X/1; X= 30/0.8 = 37.5 ~ 40 samples per fold X = 40 k = 5 # If samples per fold with 5-fold CV are too low, use 3-folds if ds.count / k < X: k = 3 # If samples are still too low, raise a value error if ds.count / k < X and not kwargs.get("permissive"): raise ValueError("Not enough samples to perform cross validation!") result = ModelParameters(cv_interval=splits['train'], cv_splits=k, task_key=kwargs.get('task_key', str(uuid4())), features=features or None) return result def _get_dataset_and_pipeline(self, model: Model, mp: ModelParameters, **kwargs): if not model.id: # Make sure the task exists model = self.model_repo.create(model) if self.model_repo.exist_parameters(model.id, mp.task_key): logging.info("Model {} Grid search {} already executed!".format( model.id, mp.task_key)) return mp # Load dataset X = self.dataset_service.get_features(model.dataset, model.symbol, mp.cv_interval.begin, mp.cv_interval.end, columns=mp.features) y = self.dataset_service.get_target(model.target, model.symbol, mp.cv_interval.begin, mp.cv_interval.end) unique, counts = np.unique(y, return_counts=True) if len(unique) < 2: logging.error( "[{}-{}-{}-{}]Training data contains less than 2 classes: {}". format(model.symbol, model.dataset, model.target, model.pipeline, unique)) raise MessageException( "Training data contains less than 2 classes: {}".format( unique)) logging.info("Dataset loaded: X {} y {} (unique: {})".format( X.shape, y.shape, unique)) # Load pipeline pipeline_module = get_pipeline(model.pipeline) return pipeline_module, X, y def grid_search(self, model: Model, mp: ModelParameters, **kwargs) -> ModelParameters: pipeline_module, X, y = self._get_dataset_and_pipeline(model, mp) tag = "{}-{}-{}-{}-{}" \ .format(model.symbol, model.dataset, model.target, model.pipeline, dict_hash(mp.parameters)) # Perform search if not kwargs.get('halving'): gscv = GridSearchCV( estimator=pipeline_module.estimator, param_grid=kwargs.get('parameter_grid', pipeline_module.PARAMETER_GRID), # cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits), cv=StratifiedKFold(n_splits=mp.cv_splits), scoring=get_precision_scorer(), verbose=kwargs.get("verbose", 0), n_jobs=kwargs.get("n_jobs", None), refit=False) else: gscv = HalvingGridSearchCV( estimator=pipeline_module.estimator, param_grid=kwargs.get('parameter_grid', pipeline_module.PARAMETER_GRID), factor=2, cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits), scoring=get_precision_scorer(), verbose=kwargs.get("verbose", 0), n_jobs=kwargs.get("n_jobs", cpu_count() / 2), refit=False, random_state=0) try: mp.start_at = get_timestamp() # Log starting timestamp gscv.fit(X, y) mp.end_at = get_timestamp() # Log ending timestamp except SplitException as e: logging.exception( "Model {} splitting yields single-class folds!\n{}".format( tag, e.message)) return mp # Fit failed, don't save this. except ValueError as e: logging.exception("Model {} raised ValueError!\n{}".format(tag, e)) return mp # Fit failed, don't save this. # Collect results results_df = pd.DataFrame(gscv.cv_results_) # Update search request with results mp.parameter_search_method = 'halving_grid_search' if kwargs.get( 'halving') else 'gridsearch' mp.parameters = gscv.best_params_ mp.cv_results = results_df.to_dict() mp.result_file = 'cv_results-{}.csv'.format(tag) # Save grid search results on storage if kwargs.get('save', True): storage_service.upload_json_obj(mp.parameters, 'grid-search-results', 'parameters-{}.json'.format(tag)) storage_service.save_df(results_df, 'grid-search-results', mp.result_file) # Update model with the new results self.model_repo.append_parameters(model.id, mp) return mp def random_search(self, model: Model, mp: ModelParameters, **kwargs) -> ModelParameters: pipeline_module, X, y = self._get_dataset_and_pipeline(model, mp) tag = "{}-{}-{}-{}-{}" \ .format(model.symbol, model.dataset, model.target, model.pipeline, dict_hash(mp.parameters)) rscv = RandomizedSearchCV(estimator=pipeline_module.estimator, param_distributions=kwargs.get( 'param_distributions', pipeline_module.PARAMETER_DISTRIBUTION), n_iter=kwargs.get('n_iter', 10), cv=StratifiedKFold(n_splits=mp.cv_splits), scoring=get_precision_scorer(), verbose=kwargs.get("verbose", 0), n_jobs=kwargs.get("n_jobs", None), refit=False, random_state=0) try: mp.start_at = get_timestamp() # Log starting timestamp rscv.fit(X, y) mp.end_at = get_timestamp() # Log ending timestamp except SplitException as e: logging.exception( "Model {} splitting yields single-class folds!\n{}".format( tag, e.message)) return mp # Fit failed, don't save this. except ValueError as e: logging.exception("Model {} raised ValueError!\n{}".format(tag, e)) return mp # Fit failed, don't save this. # Collect results results_df = pd.DataFrame(rscv.cv_results_) # Update search request with results mp.parameter_search_method = 'randomsearch' mp.parameters = rscv.best_params_ mp.result_file = 'cv_results-{}.csv'.format(tag) # Save grid search results on storage if kwargs.get('save', True): storage_service.upload_json_obj(mp.parameters, 'random-search-results', 'parameters-{}.json'.format(tag)) storage_service.save_df(results_df, 'random-search-results', mp.result_file) # Update model with the new results self.model_repo.append_parameters(model.id, mp) return mp def grid_search_new(self, symbol: str, dataset: str, target: str, pipeline: str, split: float, feature_selection_method: str, **kwargs): # Check if a model exists and has same search method existing_model = self.model_service.get_model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol) if existing_model: mp_exists = ModelService.get_model_parameters(existing_model, method='gridsearch') if mp_exists: if kwargs.get('replace'): self.model_service.remove_parameters(model=existing_model, method='gridsearch') else: if kwargs.get('save'): raise MessageException( f"Grid search already performed for {pipeline}({dataset}.{symbol}) -> {target}" ) # Retrieve dataset to use ds = self.dataset_service.get_dataset(dataset, symbol) # Determine cv_splits=K for K-fold cross validation based on dataset's sample count # Train-test split for each fold is 80% train, the lowest training window for accurate results is 30 samples # so we need X samples where X is given by the proportion: # 30/0.8 = X/1; X= 30/0.8 = 37.5 ~ 40 samples per fold X = 40 cv_splits = 5 # If samples per fold with 5-fold CV are too low, use 3-folds if ds.count / cv_splits < X: cv_splits = 3 # If samples are still too low, raise a value error if ds.count / cv_splits < X and not kwargs.get("permissive"): raise ValueError("Not enough samples to perform cross validation!") # Determine split indices based on dataset splits = DatasetService.get_train_test_split_indices(ds, split) cv_interval = splits['train'] # Load dataset features by applying a specified feature selection method X = self.dataset_service.get_dataset_features( ds=ds, begin=cv_interval['begin'], end=cv_interval['end'], method=feature_selection_method, target=target) y = self.dataset_service.get_target( name=target, symbol=symbol, begin=cv_interval['begin'], end=cv_interval['end'], ) # Check number of samples for each class in training data, if less than 3 instances are present for # each class, we're going to get a very unstable model (or no model at all for k-NN based algos) unique, counts = np.unique(y, return_counts=True) if len(unique) < 2: logging.error( "[{}-{}-{}-{}]Training data contains less than 2 classes: {}". format(symbol, dataset, target, pipeline, unique)) raise MessageException( "Training data contains less than 2 classes: {}".format( unique)) logging.info("Dataset loaded: X {} y {} (unique: {})".format( X.shape, y.shape, unique)) # Load pipeline algorithm and parameter grid pipeline_module = get_pipeline(pipeline) # Perform search gscv = GridSearchCV( estimator=pipeline_module.estimator, param_grid=kwargs.get('parameter_grid', pipeline_module.PARAMETER_GRID), # cv=BlockingTimeSeriesSplit(n_splits=mp.cv_splits), cv=StratifiedKFold(n_splits=cv_splits), scoring=get_precision_scorer(), verbose=kwargs.get("verbose", 0), n_jobs=kwargs.get("n_jobs", None), refit=False) mp = ModelParameters(cv_interval=splits['train'], cv_splits=cv_splits, task_key=kwargs.get('task_key', str(uuid4())), features=[c for c in X.columns], parameter_search_method='gridsearch') mp.start_at = get_timestamp() gscv.fit(X, y) mp.end_at = get_timestamp() # Collect results results_df = pd.DataFrame(gscv.cv_results_) mp.parameters = gscv.best_params_ mp.cv_results = results_df.loc[:, results_df.columns != 'params'].to_dict( 'records') tag = "{}-{}-{}-{}-{}".format(symbol, dataset, target, pipeline, dict_hash(mp.parameters)) mp.result_file = 'cv_results-{}.csv'.format(tag) # Is there an existing model for this search? model = Model(pipeline=pipeline, dataset=dataset, target=target, symbol=symbol, features=feature_selection_method) model.parameters.append(mp) self.model_repo.create(model) # Save grid search results on storage if kwargs.get('save', True): storage_service.upload_json_obj(mp.parameters, 'grid-search-results', 'parameters-{}.json'.format(tag)) storage_service.save_df(results_df, 'grid-search-results', mp.result_file) return mp
def model_index(query: Optional[dict] = Body(...), service: ModelService = Depends(ModelService)): return service.create_classification_models(query)
def main(dataset: str, target: str): models = ModelService() writer = pd.ExcelWriter("{}_{}.xlsx".format(dataset, target)) for symbol in symbols: model_tests = models.compare_models(symbol, dataset, target) result = [] for m in model_tests: test = m["tests"] report = test["classification_report"] if "classification_report" in test else None #duration = from_timestamp(test["end_at"]).timestamp() - from_timestamp(test["start_at"]).timestamp() # mean_dur = pd.Series(test["classification_results"]["duration"]).mean() results = ModelService.parse_test_results(test) result.append({ "pipeline": m["pipeline"], "window": test['window']['days'], #"step": str(test["step"]), "mean_fit_time": results.fit_time.mean(), "mean_predict_time": results.predict_time.mean(), "support_all": report["total_support"] if report else np.nan, "support_0": report["sup_0"] if report else np.nan, "support_1": report["sup_1"] if report else np.nan, "support_2": report["sup_2"] if report else np.nan, # Per-class precision/recall/f1/spe/geom/iba "precision_0": report["pre_0"] if report else np.nan, "recall_0": report["rec_0"] if report else np.nan, "specificity_0": report["spe_0"] if report else np.nan, "f1-score_0": report["f1_0"] if report else np.nan, "geometric_mean_0": report["geo_0"] if report else np.nan, "index_balanced_accuracy_0": report["iba_0"] if report else np.nan, "precision_1": report["pre_1"] if report else np.nan, "recall_1": report["rec_1"] if report else np.nan, "specificity_1": report["spe_1"] if report else np.nan, "f1-score_1": report["f1_1"] if report else np.nan, "geometric_mean_1": report["geo_1"] if report else np.nan, "index_balanced_accuracy_1": report["iba_1"] if report else np.nan, "precision_2": report["pre_2"] if report else np.nan, "recall_2": report["rec_2"] if report else np.nan, "specificity_2": report["spe_2"] if report else np.nan, "f1-score_2": report["f1_2"] if report else np.nan, "geometric_mean_2": report["geo_2"] if report else np.nan, "index_balanced_accuracy_2": report["iba_2"] if report else np.nan, # Roc-auc # "roc_auc_ovo_macro": report["roc_auc_ovo_macro"] if report else np.nan, # "roc_auc_ovo_weighted": report["roc_auc_ovo_weighted"] if report else np.nan, # "roc_auc_ovr_macro": report["roc_auc_ovr_macro"] if report else np.nan, # "roc_auc_ovr_weighted": report["roc_auc_ovr_weighted"] if report else np.nan, # Averages "precision_avg": report["avg_pre"] if report else np.nan, "recall_avg": report["avg_rec"] if report else np.nan, "specificity_avg": report["avg_spe"] if report else np.nan, "f1-score_avg": report["avg_f1"] if report else np.nan, "geometric_mean_avg": report["avg_geo"] if report else np.nan, "index_balanced_accuracy_avg": report["avg_iba"] if report else np.nan, }) df = pd.DataFrame(result) # Plot to XLSX with conditional formatting by coolwarm color map, # ordering by ascending accuracy df.sort_values(by='precision_avg', ascending=True)\ .style.background_gradient(cmap=cm.get_cmap('coolwarm')) \ .format(None, na_rep="-")\ .to_excel(writer, sheet_name=symbol, index_label="#", float_format = "%0.3f") # Adjust column width for column in df: column_length = max(df[column].astype(str).map(len).max(), len(column)) col_idx = df.columns.get_loc(column) + 1 writer.sheets[symbol].set_column(col_idx, col_idx, column_length) writer.close()
def main(pipeline: str, dataset: str, symbol: str, window: int): ds = DatasetService() ms = ModelService() ts = TradingService() ohlcv_ds = ds.get_dataset('ohlcv', symbol=symbol) ohlcv = ds.get_dataset_features( ohlcv_ds) # [ohlcv_ds.valid_index_min:ohlcv_ds.valid_index_max] # boll = pd.Series(percent_b(ohlcv.close, 21), index=ohlcv.index) boll = pd.Series(to_discrete_double(percent_b(ohlcv.close, 21), 20, 80), index=ohlcv.index).replace(to_replace=-1, value=np.nan) #model = ms.get_model(pipeline, dataset, 'class', symbol) _test = ms.get_test(pipeline, dataset, 'class', symbol, window) for test in [ _test ]: # I originally traded all the tests in the model. ToDo: Refactor this. # Re-convert classification results from test to a DataFrame ohlcv_results = ohlcv[test.test_interval.begin:test.test_interval.end] results = ModelService.parse_test_results(test) #results.index = ohlcv_results.index # Parse index so it's a DateTimeIndex, because Mongo stores it as a string # results.index = pd.to_datetime(results.index) asset = ts.get_asset(pipeline=pipeline, dataset=dataset, target='class', symbol=symbol, window=test.window['days']) # Now use classification results to trade! day_count = results.shape[0] cur_day = 0 print( "%B_Precision = {}", precision_score(results.label, boll.loc[results.index], average='macro', zero_division=0)) # Amount to buy in coins for buy and hold: $10k divided by first price in test set bh_price = ohlcv.close.loc[test.test_interval.begin] bh_amount = 10000 / bh_price for index, pred in results.iterrows(): cur_day += 1 # Get simulation day by converting Pandas' Timestamp to our format simulation_day = to_timestamp(index.to_pydatetime()) # Results dataframe interprets values as float, while they are actually int predicted, label = int(pred.predicted), int(pred.label) # Grab ohlcv values for current day try: values = ohlcv.loc[index] except KeyError: print(f"Day: {index} not in OHLCV index!") continue try: boll_sig = boll.loc[ index] if boll.loc[index] != np.nan else None except KeyError: boll_sig = None print(f"Day: {index} not in BOLL index!") pass _index = ohlcv.index.get_loc(index) change = TradingService.get_percent_change(values.close, values.open) print( f"Day {cur_day}/{day_count} [{index}] " f"[O {values.open} H {values.high} L {values.low} C {values.close}] " f"PCT={change}% " f"LABEL={TARGETS[label]} BPRED={TARGETS[boll_sig]} PRED={TARGETS[predicted]}" ) open_positions = ts.get_open_positions(asset=asset, day=simulation_day) for p in open_positions: p_age = TradingService.get_position_age(position=p, day=simulation_day) try: if p.type == 'MARGIN_LONG': if TradingService.check_stop_loss(p, values.low): ts.close_long(asset=asset, day=simulation_day, close_price=p.stop_loss, position=p, detail='Stop Loss') elif TradingService.check_take_profit(p, values.high): ts.close_long(asset=asset, day=simulation_day, close_price=p.take_profit, position=p, detail='Take Profit') elif predicted == SELL: ts.close_long(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Sell Signal') elif predicted == HOLD and p_age > 86400 * 3: ts.close_long(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Age') elif predicted == BUY: if change > 0: ts.update_stop_loss(asset=asset, position=p, close_price=values.close, pct=-0.05) elif p.type == 'MARGIN_SHORT': if TradingService.check_stop_loss(p, values.high): ts.close_short(asset=asset, day=simulation_day, close_price=p.stop_loss, position=p, detail='Stop Loss') elif TradingService.check_take_profit(p, values.low): ts.close_short(asset=asset, day=simulation_day, close_price=p.take_profit, position=p, detail='Take Profit') elif predicted == SELL: # If we had some profit and signal is still SELL, book those by lowering stop loss if change < 0: ts.update_stop_loss(asset=asset, position=p, close_price=values.close, pct=0.05) elif predicted == HOLD and p_age > 86400 * 3: ts.close_short(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Age') elif predicted == BUY: ts.close_short(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Buy Signal') except MessageException as e: print(f"Order handling exception: {e.message}") try: # If prediction is BUY (price will rise) then open a MARGIN LONG position if predicted == BUY: ts.open_long(asset=asset, day=simulation_day, close_price=values.close, size=0.1, stop_loss=-0.1, take_profit=0.05) # If prediction is SELL (price will drop) open a MARGIN SHORT position elif predicted == SELL: ts.open_short(asset=asset, day=simulation_day, close_price=values.close, size=0.1, stop_loss=0.1, take_profit=-0.05) except MessageException as e: print(f"Order placement exception: {e.message}") # If this is the last trading day of the period, close all open positions if index.timestamp() == results.index[-1].timestamp(): print("Last trading day reached, liquidating all positions..") open_positions = ts.get_open_positions(asset=asset, day=simulation_day) for p in open_positions: try: if p.type == 'MARGIN_LONG': ts.close_long(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Liquidation') elif p.type == 'MARGIN_SHORT': ts.close_short(asset=asset, day=simulation_day, close_price=values.close, position=p, detail='Liquidation') except MessageException as e: print(f"Order liquidation exception: {e.message}") # Update equity value for the asset ts.update_equity(asset=asset, day=simulation_day, price=values.close) # Update baseline values for the asset ts.update_baseline(asset=asset, day=simulation_day, name='buy_and_hold', value=values.close * bh_amount) print("Timeframe done.")