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 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(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(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 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(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(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")