def get(self, use_db, model_type, symbol, unit, count, period, partition,
delta, seed, trees, jobs, longs):
# If using db then validate db
if use_db:
@validate_db(db)
def get_queries():
return [query_to_dict(q) for q in Data.query.all()]
json = get_queries()
else:
json = None
m = market.Market(json=json,
symbol=symbol,
unit=unit,
count=count,
period=period)
features = m.set_features(partition=partition)
features = m.set_long_features(features,
columns_to_set=longs,
partition=partition)
targets = market.set_targets(features, delta=delta)
features = features.drop(['close'], axis=1)
model = market.setup_model(features[:-1],
targets,
model_type=model_type.lower(),
seed=seed,
n_estimators=trees,
n_jobs=jobs)
next_date = features.tail(
1) # Remember the entry we didn't train? Predict it.
trend = market.target_code_to_name(model._predict_trends(next_date)[0])
accuracy = model.accuracy(model.features.test, model.targets.test)
proba = model._predict_probas(next_date)
proba_log = model._predict_logs(next_date) # Logarithmic scale
return {
'trend': trend,
'test_set_accuracy': accuracy,
'probabilities': {
market.target_code_to_name(code): p
for code, p in enumerate(proba[0])
}
}
def main():
args = get_args()
assert args.partition > 0, 'The data must be partitioned!'
m = market.Market(symbol=args.symbol,
unit=args.unit,
count=args.count,
period=args.period)
x = m.features(partition=args.partition)
if args.long is not None:
# Create long features DataFrame
x_long = m.features(partition=2 * args.partition)
# Remove features not specified by args.long
unwanted_features = [f for f in x.columns if f not in args.long]
x_long = x_long.drop(unwanted_features, axis=1)
# Prefix long columns with 'long_' to fix naming conflicts
x_long.columns = ['long_{0}'.format(f) for f in x_long.columns]
# Merge the two DataFrames
skip = args.partition
x = pd.concat([x[skip:].reset_index(drop=True), x_long], axis=1)
y = market.targets(x, delta=args.delta)
x = x.drop(['close'], axis=1)
model = market.setup_model(x[:-1],
y,
model_type=args.model.lower(),
seed=args.seed,
n_estimators=args.trees,
n_jobs=args.jobs)
next_date = x.tail(1) # Remember the entry we didn't train? Predict it.
# TODO: Reimplement display of confusion matrix and feature importances
acc = model.accuracy(model.features.test, model.targets.test)
print('Test Set Accuracy: {0:.3f}%'.format(100 * acc))
trends = model._predict_trends(next_date)
print('Predicted Trend: {0}'.format(market.target_code_to_name(trends[0])))
if args.proba:
probas = model._predict_probas(next_date)
print('Probability: {0}'.format(probas[0]))
if args.proba_log:
logs = model._predict_logs(next_date)
print('Log Probability: {0}'.format(logs[0]))
def get(self, use_db, model, symbol, unit, count, period, partition, delta,
seed, trees, jobs, longs):
if use_db:
db_json = None
else:
db_json = None
m = market.Market(json=db_json,
symbol=symbol,
unit=unit,
count=count,
period=period)
features = m.set_features(partition=partition)
features = m.set_long_features(features,
columns_to_set=longs,
partition=partition)
targets = market.set_targets(features, delta=delta)
features = features.drop(['close'], axis=1)
model = market.setup_model(features[:-1],
targets,
model_type=model.lower(),
seed=seed,
n_estimators=trees,
n_jobs=jobs)
next_date = features.tail(
1) # Remember the entry we didn't train? Predict it.
trend = market.target_code_to_name(model._predict_trends(next_date)[0])
accuracy = model.accuracy(model.features.test, model.targets.test)
proba = model._predict_probas(next_date)
proba_log = model._predict_logs(next_date) # Logarithmic scale
return {
"trend": trend,
"test_set_accuracy": accuracy,
"probabilities": {
market.target_code_to_name(code): p
for code, p in enumerate(proba[0])
}
}
def predict_coin(self, coin, unit, api, model_type):
if api == 'spec':
pcoin = 'USDT_' + coin.strip()
coinmarket = market.Market(symbol=pcoin,
unit=unit,
count=6,
period=86400)
#Retrieve the x and y axises
x = coinmarket.features(partition=14)
y = market.targets(x, delta=25)
#Get rid of the close stat because it is not useful at all.
x = x.drop(['close'], axis=1)
#Now create the random forest model or other one supported by the package
model = market.setup_model(x[:-1],
y,
model_type=model_type,
seed=1,
n_estimators=65,
n_jobs=4)
# Predict the target test set from the features test set
trends = model._predict_trends(model.features.test)
# Get accuracies
ftr_imps = model.feature_importances()
conf_mx = model.confusion_matrix(model.targets.test, trends)
acc = model.accuracy(model.features.test, model.targets.test)
#Predictions and probabilities for the next trend
next_date = x.tail(1)
trends = model._predict_trends(next_date)
pt = "The market for {} is predicted to be {}!".format(
coin, market.target_code_to_name(trends[0]))
probas = model._predict_probas(next_date)
pr = 'Probability: {0}'.format(probas[0])
print(pr)
logs = model._predict_logs(next_date)
print('Log Probability: {0}'.format(logs[0]))
return next_date, pt
def main():
args = get_args()
assert args.partition > 0, 'The data must be partitioned!'
m = market.Market(symbol=args.symbol, unit=args.unit,
count=args.count, period=args.period)
features = m.set_features(partition=args.partition)
if args.long is not None:
features = m.set_long_features(features,
columns_to_set=args.long,
partition=args.partition)
targets = market.set_targets(features, delta=args.delta)
features = features.drop(['close'], axis=1)
model = market.setup_model(features[:-1], targets,
model_type=args.model.lower(),
seed=args.seed,
n_estimators=args.trees,
n_jobs=args.jobs)
next_date = features.tail(1) # Remember the entry we didn't train? Predict it.
# TODO: Reimplement display of confusion matrix and feature importances
acc = model.accuracy(model.features.test, model.targets.test)
print('Test Set Accuracy: {0:.3f}%'.format(100 * acc))
trends = model._predict_trends(next_date)
print('Predicted Trend: {0}'.format(market.target_code_to_name(trends[0])))
if args.proba:
probas = model._predict_probas(next_date)
print('Probability: {0}'.format(probas[0]))
if args.proba_log:
logs = model._predict_logs(next_date)
print('Log Probability: {0}'.format(logs[0]))
from speculator import market
# Init market raw data
m = market.Market(symbol='USDT_BTC', unit='month', count=6, period=86400)
# Parse features, x axis
x = m.features(partition=14)
# Parse targets, y axis
y = market.targets(x, delta=25)
# Create the random forest model
# The last entry doesn't have a target (can't predict yet), so skip over it
model = market.setup_model(x[:-1],
y,
model_type='random_forest',
seed=1,
n_estimators=65,
n_jobs=4)
# Predict the target test set from the features test set
pred = model.predict(model.features.test)
# Get accuracies
ftr_imps = model.feature_importances()
conf_mx = model.confusion_matrix(model.targets.test, pred)
acc = model.accuracy(model.targets.test, pred)
# Display accuracies
print('##################')
print('# TEST SET #')
print('##################')
def main():
args = get_args()
m = market.Market(symbol=args.symbol,
unit=args.unit,
count=args.count,
period=args.period)
x = m.features(partition=args.partition)
if args.long is not None:
# Create long features DataFrame
x_long = m.features(partition=2 * args.partition)
# Remove features not specified by args.long
unwanted_features = [f for f in x.columns if f not in args.long]
x_long = x_long.drop(unwanted_features, axis=1)
# Prefix long columns with 'long_' to fix naming conflicts
x_long.columns = ['long_{0}'.format(f) for f in x_long.columns]
# Merge the two DataFrames
skip = args.partition
x = pd.concat([x[skip:].reset_index(drop=True), x_long], axis=1)
y = market.targets(x, delta=args.delta)
x = x.drop(['close'], axis=1)
model = market.setup_model(x[:-1],
y,
model_type=args.model,
seed=args.seed,
n_estimators=args.trees,
n_jobs=args.jobs)
# Predict the target test set from the features test set
pred = model.predict(model.features.test)
# Get accuracies
ftr_imps = model.feature_importances()
conf_mx = model.confusion_matrix(model.targets.test, pred)
acc = model.accuracy(model.targets.test, pred)
# Display accuracies
print('##################')
print('# TEST SET #')
print('##################')
print('Accuracy: {0:.3f}%'.format(100 * acc))
print('\nConfusion Matrix:')
print(conf_mx)
print(market.TARGET_CODES)
print('\nFeature Importance:')
for ftr, imp in ftr_imps:
print(' {0}: {1:.3f}%'.format(ftr, 100 * imp))
print()
# Display prediction and probabilities for the next trend
print('##################')
print('# PREDICTED NEXT #')
print('##################')
next_date = x.tail(1) # Remember the entry we didn't train? Predict it.
trend = market.target_code_to_name(model.predict(next_date)[0])
print('Trend: {0}'.format(trend))
if args.proba:
print('Probability: {0}'.format(model.predict_proba(next_date)))
if args.proba_log:
print('Log Probability: {0}'.format(
model.predict_log_proba(next_date)))