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 predict_day(self, pipeline: str, dataset: str, target: str,
symbol: str, day: str, window: dict):
model = self.get_model(pipeline, dataset, target, symbol)
# Load dataset
ds = DatasetService()
d = ds.get_dataset(model.dataset, model.symbol)
# Get training data including the first training window
begin = sub_interval(timestamp=day, interval=window)
if from_timestamp(d.valid_index_min).timestamp() > from_timestamp(
begin).timestamp():
raise MessageException("Not enough data for training! [Pipeline: {} Dataset: {} Symbol: {} Window: {}]" \
.format(model.pipeline, model.dataset, model.symbol, window))
X = ds.get_features(model.dataset, model.symbol, begin=begin, end=day)
y = ds.get_target(model.target, model.symbol, begin=begin, end=day)
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))
# Load pipeline
pipeline_module = get_pipeline(model.pipeline)
# Slice testing interval in windows
df = predict_day(pipeline_module.estimator, model.parameters[-1], X, y,
day)
return df
def create_model_test(self,
*,
model: Model,
split=0.7,
step=None,
task_key=None,
window=None,
**kwargs):
service = DatasetService()
ds = service.get_dataset(model.dataset, model.symbol)
splits = DatasetService.get_train_test_split_indices(ds, split)
parameters = kwargs.get('parameters')
features = kwargs.get('features')
if isinstance(parameters, str) and parameters == 'latest':
if model.parameters:
parameters = model.parameters[-1].parameters
else:
parameters = None
if isinstance(features, str):
fs = DatasetService.get_feature_selection(ds=ds,
method=features,
target=model.target)
if fs:
features = fs.features
else:
features = None
result = ModelTest(window=window or {'days': 30},
step=step or ds.interval,
parameters=parameters or {},
features=features or [],
test_interval=splits['test'],
task_key=task_key or str(uuid4()))
return result
def get_dataset(
symbol: Optional[str] = None,
service: DatasetService = Depends(DatasetService),
):
if symbol:
return [d for d in service.find_by_symbol(symbol)]
else:
return [d for d in service.all()]
def build(symbol: str, builder: str, ohlcv: str, coinmetrics: str):
build_args = {'ohlcv': ohlcv, 'coinmetrics': coinmetrics}
print("Building {} [{} -> {}]".format(symbol, build_args, builder))
bs = DatasetBuildingService()
bs.check_builder_args(builder, build_args)
df = bs.build_dataset(symbol, builder, build_args)
ds = DatasetService()
ds.create_dataset(df, builder, symbol,
'FEATURES' if builder != 'target' else 'TARGET')
print("Done")
def test_model(self, model: Model, mt: ModelTest, **kwargs):
if not model.id:
model = self.model_repo.create(model)
if self.model_repo.exist_test(model.id, mt.task_key):
logging.info("Model {} test {} already executed!".format(
model.id, mt.task_key))
return mt
# Load dataset
ds = DatasetService()
d = ds.get_dataset(model.dataset, model.symbol)
# Get training data including the first training window
begin = sub_interval(timestamp=mt.test_interval.begin,
interval=mt.window)
end = add_interval(timestamp=mt.test_interval.end, interval=mt.step)
if from_timestamp(d.valid_index_min).timestamp() > from_timestamp(
begin).timestamp():
raise MessageException("Not enough data for training! [Pipeline: {} Dataset: {} Symbol: {} Window: {}]" \
.format(model.pipeline, model.dataset, model.symbol, mt.window))
X = ds.get_features(model.dataset, model.symbol, begin=begin, end=end)
y = ds.get_target(model.target, model.symbol, begin=begin, end=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))
# Load pipeline
pipeline_module = get_pipeline(model.pipeline)
# Slice testing interval in windows
ranges = timestamp_windows(begin, end, mt.window, mt.step)
mt.start_at = get_timestamp()
df = test_windows(pipeline_module.estimator, mt.parameters, X, y,
ranges)
mt.end_at = get_timestamp()
mt.classification_results = df.to_dict()
clf_report = flattened_classification_report_imbalanced(
df.label, df.predicted)
roc_report = roc_auc_report(
df.label, df.predicted,
df[[c for c in df.columns if '_proba_' in c]])
clf_report.update(roc_report)
mt.classification_report = clf_report
self.model_repo.append_test(model.id, mt)
return mt
def get_dataset_csv(
dataset_id: str,
service: DatasetService = Depends(DatasetService),
):
ds = service.get(dataset_id)
df = service.get_features(name=ds.name,
symbol=ds.symbol,
begin=ds.index_min,
end=ds.index_max,
columns=ds.features)
return df.to_csv(index_label='time')
def main(dataset: str):
ds_service = DatasetService()
symbols = ds_service.get_dataset_symbols(name=dataset)
ds_data = {
s: ds_service.get_dataset(name=dataset, symbol=s).features
for s in symbols
}
# We need to reshape / flatten data
records = []
symbol_lookup = {s: i for i, s in enumerate(symbols)}
for symbol, features in ds_data.items():
record = {
'symbol': symbol.replace('USD', ''),
#'symbol_id': symbol_lookup[symbol]
}
for f in features:
if f.startswith('adrbal1in') and f.endswith('cnt'):
f = 'adrbal1in{N}cnt'
elif f.startswith('adrbalntv') and f.endswith('cnt'):
f = 'adrbalntv{N}cnt'
elif f.startswith('splyact') and not 'pct' in f:
f = 'splyact{T}'
elif f.startswith('splyadrbal1in'):
f = 'splyadrbal1in{N}'
elif f.startswith('splyadrbalntv'):
f = 'splyadrbalntv{N}'
elif f.startswith('splyadrtop'):
f = 'splyadrtop{N}'
elif f.startswith('adrbalusd') and f.endswith('cnt'):
f = 'adrbalusd{N}cnt'
elif f.startswith('splyadrbalusd'):
f = 'splyadrbalusd{N}'
elif f.startswith('txtfrval') and f.endswith('ntv'):
f = 'txtfrval{A}ntv'
elif f.startswith('txtfrval') and f.endswith('usd'):
f = 'txtfrval{A}usd'
elif f.startswith('fee') and f.endswith('usd'):
f = 'fee{A}usd'
elif f.startswith('gaslmtblk'):
f = 'gaslmtblk'
elif f.startswith('gaslmttx'):
f = 'gaslmttx'
elif f.startswith('gasusedtx'):
f = 'gasusedtx'
elif f.startswith('isccont'):
f = 'isscont'
record[f] = 'Y'
records.append(record)
result_frame = pd.DataFrame.from_records(records).fillna(value='N')
#result_frame.set_index(keys='symbol', inplace=True)
result_frame = result_frame.set_index(keys='symbol').T
latex = result_frame.to_latex()
print(result_frame.head())
def get_dataset(
symbol: str,
dataset: Optional[str] = None,
target: Optional[str] = None,
begin: Optional[str] = None,
end: Optional[str] = None,
service: DatasetService = Depends(DatasetService),
):
if not dataset and not target:
raise HTTPException(
status_code=400,
detail=
"At least one of 'dataset' or 'target' parameters must be specified!"
)
_name = dataset
if not _name:
_name = 'target'
d = service.get_dataset(name=_name, symbol=symbol)
# If begin/end not specified, use recorded.
# If auto use valid.
if not begin:
begin = d.index_min
elif begin == 'auto':
begin = d.valid_index_min
if not end:
end = d.index_max
elif end == 'auto':
end = d.valid_index_max
# Retrieve dataframes
dfs = []
if dataset:
df = service.get_features(name=dataset,
symbol=symbol,
begin=begin,
end=end)
dfs.append(df)
if target:
dfs.append(
service.get_target(name=target,
symbol=symbol,
begin=begin,
end=end))
# Concatenate dataframes and target
res = pd.concat(dfs, axis='columns') if len(dfs) > 1 else dfs[0]
# Return CSV
return res.to_csv(index_label='time')
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 main(dataset: str, target: str, symbol: str):
ds_service = DatasetService()
ds = ds_service.get_dataset(name=dataset, symbol=symbol)
fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target='class')
# hierarchy = load_hierarchy(f"{dataset}_{target}_feature_hierarchy.yml", importances=fs.feature_importances)
# hdf = pd.DataFrame(hierarchy)
# fig = px.treemap(hdf, path=['category', 'subgroup', 'name'], values='importance')
# fig.show()
#
# fig = px.sunburst(hdf, path=['category', 'subgroup', 'name'], values='importance')
# fig.show()
shap_values, shap_expected_values = parse_shap_values(fs.shap_values)
X = ds_service.get_dataset_features(ds=ds, begin=fs.search_interval.begin, end=fs.search_interval.end)
y = ds_service.get_target(name='class', symbol=symbol, begin=fs.search_interval.begin, end=fs.search_interval.end)
fig = plt.figure()
plt.suptitle(f"Shap summary plot for {dataset}.{symbol} -> {target}")
shap.summary_plot(shap_values, X, class_names=["SELL", "HOLD", "BUY"], show=False, max_display=352, use_log_scale=True)
plt.tight_layout()
fig.show()
shap_dfs = []
for cls, arr in enumerate(shap_values):
class_df = pd.DataFrame(arr, columns=X.columns, index=X.index)
class_df.columns = [f"{c}_class{cls}" for c in class_df.columns]
shap_dfs.append(class_df)
shap_df = pd.concat(shap_dfs, axis='columns')
shap_df = shap_df.reindex(sorted(shap_df.columns), axis=1)
print(shap_df.head())
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 main(dataset: str,
target: str,
splits: Optional[int] = 1,
type: Optional[str] = 'sh'):
dss = DatasetService()
symbols = dss.get_dataset_symbols(name=dataset)
lines = []
for symbol in symbols:
for pipeline in PIPELINE_LIST:
lines.append(
f"python grid_search_new.py {symbol} {dataset} {target} {pipeline} --feature-selection-method importances_shap"
)
destfile = f"gridsearch_{dataset}_{target}_all"
if type == 'cmd':
with open(destfile + ".cmd", "w") as f:
f.write("\n".join(["@echo off"] + lines))
elif type == 'sh':
with open(destfile + ".sh", "w") as f:
f.write("\n".join(["#!/bin/bash"] + lines))
print(f"Grid search script saved to {destfile}")
return destfile
def main(dataset: str, target: str, symbol: str):
ds_service = DatasetService()
ds = ds_service.get_dataset(name=dataset, symbol=symbol)
fs = DatasetService.get_feature_selection(ds=ds, method='importances_shap', target='class')
# shap_values, shap_expected_values = parse_shap_values(fs.shap_values)
# X = ds_service.get_dataset_features(ds=ds, begin=fs.search_interval.begin, end=fs.search_interval.end)
# shap_df_0 = pd.DataFrame(data=shap_values[0], index=X.index, columns=X.columns)
# shap_df_1 = pd.DataFrame(data=shap_values[1], index=X.index, columns=X.columns)
# shap_df_2 = pd.DataFrame(data=shap_values[2], index=X.index, columns=X.columns)
hierarchy = load_hierarchy(f"{dataset}_{target}_feature_hierarchy.yml", importances=fs.feature_importances)
# for record in hierarchy:
# feature = record['name']
# try:
# record['shap_mean_0'] = shap_df_0[feature].mean()
# record['shap_mean_1'] = shap_df_1[feature].mean()
# record['shap_mean_2'] = shap_df_2[feature].mean()
# except KeyError as e:
# print(f"Feature {feature} not in dataset!")
# record['shap_mean_0'] = np.nan
# record['shap_mean_1'] = np.nan
# record['shap_mean_2'] = np.nan
# pass
os.makedirs(f"data/selection_{dataset}_{target}/", exist_ok=True)
hdf = pd.DataFrame(hierarchy)
csv_name = f"data/selection_{dataset}_{target}/{symbol}_feature_importances.csv"
hdf.to_csv(csv_name, index_label='index')
print(f"Augmented importances dataframe exported to {csv_name}")
csv_name = f"data/selection_{dataset}_{target}/{symbol}_feature_importances_selected.csv"
hdf[hdf.name.isin(fs.features)].to_csv(csv_name, index_label='index')
print(f"Augmented selected features dataframe exported to {csv_name}")
def create_classification_models(self, query, pipeline):
ds = DatasetService()
models = []
if query is None:
query = {{
"type": "FEATURES",
}}
datasets = ds.query(query)
# All possible combinations
all_models = {}
for d in datasets:
# Get targets for this symbol
tgt = ds.get_dataset('target', d.symbol)
if not d.symbol in all_models:
all_models[d.symbol] = []
for t, p in itertools.product(tgt.features, PIPELINE_LIST):
if t in ['price', 'pct']:
continue
all_models[d.symbol].append((d, t, p))
# Method to process a batch of items
results = Parallel(n_jobs=-1)(
delayed(create_models_batch)(symbol, items)
for symbol, items in all_models.items())
return [item for sublist in results for item in sublist]
def _import(ticket: ImportRequest = Body(...),
service: DatasetService = Depends(DatasetService)):
filename = secure_filename(ticket.name)
if not ticket.dataset:
raise HTTPException(status_code=400,
detail='You must complete the dataset field!')
# Import dataset to feature repository
try:
ds = service.import_from_storage(ticket.bucket, filename,
ticket.dataset, ticket.symbol)
return ds.dict()
except Exception as e:
logging.exception(e)
raise HTTPException(status_code=500,
detail='Failed to import data in repository')
def merge_dataset(name: str,
symbol: str,
sync: Optional[bool] = False,
query: dict = Body(...),
service: DatasetService = Depends(DatasetService),
repo: DatasetRepository = Depends(DatasetRepository),
tasks: TaskService = Depends(TaskService)):
if sync:
datasets = repo.query(query)
return service.merge_datasets(datasets, name, symbol)
else:
r = MergeRequest(query=query, name=name, symbol=symbol)
return tasks.send(task_name='merge_datasets',
task_args=r.dict(),
name='merge_datasets-{}->{}-{}'.format(
str(r.query), r.name, r.symbol),
batch=str(r.query))
def create_features_search(self,
*,
symbol: str,
dataset: str,
target: str,
split: float,
method: str,
task_key: str = None) -> ModelFeatures:
ds = self.dataset_service.get_dataset(dataset, symbol)
splits = DatasetService.get_train_test_split_indices(ds, split)
result = ModelFeatures(dataset=dataset,
target=target,
symbol=symbol,
search_interval=splits['train'],
feature_selection_method=method,
task_key=task_key or str(uuid4()))
return result
def _import_many(tickets: List[ImportRequest] = Body(...),
service: DatasetService = Depends(DatasetService)):
results = []
for ticket in tickets:
filename = secure_filename(ticket.name)
if not ticket.dataset:
logging.error("Skipping {}: missing dataset!".format(
ticket.symbol))
continue
# Import dataset to feature repository
try:
logging.info("Importing: {}".format(ticket.dict()))
ds = service.import_from_storage(ticket.bucket, filename,
ticket.dataset, ticket.symbol)
results.append(ds.dict())
except Exception as e:
logging.exception(e)
continue
return results
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 task_build_dataset(req: dict):
req = MergeRequest(**req)
service: DatasetService = DatasetService()
datasets = service.query(req.query)
ds = service.merge_datasets(datasets, req.name, req.symbol)
return ds.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)
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.")
def __init__(self):
self.model_repo = ModelRepository()
self.model_service = ModelService()
self.dataset_service = DatasetService()
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 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 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, 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}")
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 main(dataset: str):
dss = DatasetService()
records = []
for symbol in SYMBOLS:
ds = dss.get_dataset(name=dataset, symbol=symbol)
fs = DatasetService.get_feature_selection(ds, 'importances_shap',
'class')
target = dss.get_dataset_target(ds=ds, name='class')
uniq, cnt = np.unique(target, return_counts=True)
if cnt[0] + cnt[1] + cnt[2] != ds.count:
print(f"Mismatch between classes and count in {symbol}")
mindt = from_timestamp(ds.valid_index_min)
maxdt = from_timestamp(ds.valid_index_max)
daysn = (maxdt - mindt).days
records.append({
'Pair': symbol,
'num_features': len(ds.features),
'sel_features': len(fs.features),
'min_index': ds.valid_index_min,
'max_index': ds.valid_index_max,
'valid_days': daysn,
'records': ds.count,
'sell_count': cnt[0],
'hold_count': cnt[1],
'buy_count': cnt[2]
})
df = pd.DataFrame.from_records(records)
fig = px.timeline(df, x_start="min_index", x_end="max_index", y="Pair")
fig.update_yaxes(
autorange="reversed") # otherwise tasks are listed from the bottom up
#fig.show()
fig.update_layout(title={
'text': f"Sample distribution across datasets",
'y': 0.9,
'x': 0.5,
'xanchor': 'center',
'yanchor': 'top'
},
paper_bgcolor='rgba(0,0,0,0)',
plot_bgcolor='rgba(0,0,0,0.5)',
font={'color': 'White'},
margin={
'l': 5,
'r': 5,
't': 80,
'b': 5,
'pad': 5
})
fig.write_image("images/data_summary/timeline.png")
for symbol in SYMBOLS:
sdf = df[df.Pair == symbol]
pie_values = [
sdf['sell_count'].values[0], sdf['hold_count'].values[0],
sdf['buy_count'].values[0]
]
pie_labels = ['SELL', 'HOLD', 'BUY']
sfig = go.Figure(data=[
go.Pie(
labels=pie_labels,
values=pie_values,
textinfo='label+percent',
#insidetextorientation='radial',
showlegend=False)
])
sfig.update_layout(title={
'text': f"Class distribution for pair {symbol}",
'y': 0.9,
'x': 0.5,
'xanchor': 'center',
'yanchor': 'top',
'font': {
'size': 22
}
},
paper_bgcolor='rgba(0,0,0,0)',
plot_bgcolor='rgba(0,0,0,0)',
font={
'color': 'White',
'size': 26
},
margin={
'l': 0,
'r': 0,
't': 80,
'b': 0,
'pad': 0
},
uniformtext_minsize=24)
sfig.write_image(f"images/data_summary/{symbol}_distribution.png")
print(df.head())
def task_build_dataset(req: dict):
req = BuildRequest(**req)
service: DatasetService = DatasetService()
ds = service.build_dataset(req.symbol, req.builder, req.args)
return ds.dict()