Building Minute-Level Price and Volume Features in Qlib
Code Qlib
Summary
The document defines Qlib data handlers for one-minute bars, aimed at high-frequency research and backtesting. The training handler creates normalized open, high, low, close, and approximate VWAP features, along with volume features. Price features are scaled against the last value of the trading day and include a lagged version; volume is normalized against a rolling mean. Paused observations are masked, missing values are filled, and suspicious volume bars are filtered using the estimated VWAP’s position relative to the bar’s high and low.
A separate backtest handler supplies close, approximate VWAP, and filtered volume inputs without the same normalization and processor setup. The code comments explain that the source data lacks VWAP, so it estimates it with a weighted combination of OHLC prices. These are feature-construction choices, not evidence of predictive value: the excerpt reports no tests, performance results, or validation of the approximation and filters.
Key ideas
- The handlers load one-minute data and construct inputs for model training and backtesting.
- Price features are normalized to the trading day’s final observed price, with additional lagged features.
- VWAP is approximated from open, high, low, and close because the data source lacks a VWAP field.
- The feature logic masks paused observations, fills missing values, and filters some volume bars.
- The document does not report predictive tests or trading performance for these features.
Tags
Full text
# highfreq_handler.py
```py
from qlib.data.dataset.handler import DataHandler, DataHandlerLP
from qlib.contrib.data.handler import check_transform_proc
class HighFreqHandler(DataHandlerLP):
def __init__(
self,
instruments="csi300",
start_time=None,
end_time=None,
infer_processors=[],
learn_processors=[],
fit_start_time=None,
fit_end_time=None,
drop_raw=True,
):
infer_processors = check_transform_proc(infer_processors, fit_start_time, fit_end_time)
learn_processors = check_transform_proc(learn_processors, fit_start_time, fit_end_time)
data_loader = {
"class": "QlibDataLoader",
"kwargs": {
"config": self.get_feature_config(),
"swap_level": False,
"freq": "1min",
},
}
super().__init__(
instruments=instruments,
start_time=start_time,
end_time=end_time,
data_loader=data_loader,
infer_processors=infer_processors,
learn_processors=learn_processors,
drop_raw=drop_raw,
)
def get_feature_config(self):
fields = []
names = []
template_if = "If(IsNull({1}), {0}, {1})"
template_paused = "Select(Or(IsNull($paused), Eq($paused, 0.0)), {0})"
template_fillnan = "BFillNan(FFillNan({0}))"
# Because there is no vwap field in the yahoo data, a method similar to Simpson integration is used to approximate vwap
simpson_vwap = "($open + 2*$high + 2*$low + $close)/6"
def get_normalized_price_feature(price_field, shift=0):
"""Get normalized price feature ops"""
if shift == 0:
template_norm = "Cut({0}/Ref(DayLast({1}), 240), 240, None)"
else:
template_norm = "Cut(Ref({0}, " + str(shift) + ")/Ref(DayLast({1}), 240), 240, None)"
feature_ops = template_norm.format(
template_if.format(
template_fillnan.format(template_paused.format("$close")),
template_paused.format(price_field),
),
template_fillnan.format(template_paused.format("$close")),
)
return feature_ops
fields += [get_normalized_price_feature("$open", 0)]
fields += [get_normalized_price_feature("$high", 0)]
fields += [get_normalized_price_feature("$low", 0)]
fields += [get_normalized_price_feature("$close", 0)]
fields += [get_normalized_price_feature(simpson_vwap, 0)]
names += ["$open", "$high", "$low", "$close", "$vwap"]
fields += [get_normalized_price_feature("$open", 240)]
fields += [get_normalized_price_feature("$high", 240)]
fields += [get_normalized_price_feature("$low", 240)]
fields += [get_normalized_price_feature("$close", 240)]
fields += [get_normalized_price_feature(simpson_vwap, 240)]
names += ["$open_1", "$high_1", "$low_1", "$close_1", "$vwap_1"]
fields += [
"Cut({0}/Ref(DayLast(Mean({0}, 7200)), 240), 240, None)".format(
"If(IsNull({0}), 0, If(Or(Gt({1}, Mul(1.001, {3})), Lt({1}, Mul(0.999, {2}))), 0, {0}))".format(
template_paused.format("$volume"),
template_paused.format(simpson_vwap),
template_paused.format("$low"),
template_paused.format("$high"),
)
)
]
names += ["$volume"]
fields += [
"Cut(Ref({0}, 240)/Ref(DayLast(Mean({0}, 7200)), 240), 240, None)".format(
"If(IsNull({0}), 0, If(Or(Gt({1}, Mul(1.001, {3})), Lt({1}, Mul(0.999, {2}))), 0, {0}))".format(
template_paused.format("$volume"),
template_paused.format(simpson_vwap),
template_paused.format("$low"),
template_paused.format("$high"),
)
)
]
names += ["$volume_1"]
return fields, names
class HighFreqBacktestHandler(DataHandler):
def __init__(
self,
instruments="csi300",
start_time=None,
end_time=None,
):
data_loader = {
"class": "QlibDataLoader",
"kwargs": {
"config": self.get_feature_config(),
"swap_level": False,
"freq": "1min",
},
}
super().__init__(
instruments=instruments,
start_time=start_time,
end_time=end_time,
data_loader=data_loader,
)
def get_feature_config(self):
fields = []
names = []
template_if = "If(IsNull({1}), {0}, {1})"
template_paused = "Select(Or(IsNull($paused), Eq($paused, 0.0)), {0})"
template_fillnan = "BFillNan(FFillNan({0}))"
# Because there is no vwap field in the yahoo data, a method similar to Simpson integration is used to approximate vwap
simpson_vwap = "($open + 2*$high + 2*$low + $close)/6"
fields += [
"Cut({0}, 240, None)".format(template_fillnan.format(template_paused.format("$close"))),
]
names += ["$close0"]
fields += [
"Cut({0}, 240, None)".format(
template_if.format(
template_fillnan.format(template_paused.format("$close")),
template_paused.format(simpson_vwap),
)
)
]
names += ["$vwap0"]
fields += [
"Cut(If(IsNull({0}), 0, If(Or(Gt({1}, Mul(1.001, {3})), Lt({1}, Mul(0.999, {2}))), 0, {0})), 240, None)".format(
template_paused.format("$volume"),
template_paused.format(simpson_vwap),
template_paused.format("$low"),
template_paused.format("$high"),
)
]
names += ["$volume0"]
return fields, names
```Shown in full with attribution under the source's licence. Licence: MIT
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.