Support Highfreq Backtest with the Model/Rule/RL Strategy #408

bxdd · 2021-04-30T14:59:44Z

Description

Support Highfreq Backtest with the Model/Rule/RL Strategy

Motivation and Context

How Has This Been Tested?

Pass the test by running: pytest qlib/tests/test_all_pipeline.py under upper directory of qlib.
If you are adding a new feature, test on your own test scripts.

Screenshots of Test Results (if appropriate):

Pipeline test:
Your own tests:

Types of changes

Fix bugs
Add new feature
Update documentation

you-n-g · 2021-05-07T04:06:35Z

qlib/contrib/backtest/__init__.py

@@ -2,12 +2,12 @@
 # Licensed under the MIT License.

 from .order import Order


You could move the core framework outside of contrib folder now.

You can move them later

you-n-g · 2021-05-07T04:17:59Z

examples/highfreq/backtest/workflow.py

@@ -0,0 +1,145 @@
+#  Copyright (c) Microsoft Corporation.


Our framework is a Multi-level Trading system instead of a single high frequency trading system.
The folder name could have a better name than highfreq.

you-n-g · 2021-05-07T04:26:51Z

examples/highfreq/backtest/workflow.py

+            },
+        },
+        "backtest": {
+            "start_time": trade_start_time,


Part of the config looks like belonging to env.

you-n-g · 2021-05-07T04:39:25Z

qlib/utils/__init__.py

+
+
+def parse_freq(freq):
+    freq = freq.lower()


you-n-g · 2021-05-07T06:54:06Z

qlib/workflow/record_temp.py

+            ret_freq.extend(self._get_report_freq(env_config["kwargs"]["sub_env"]))
+        return ret_freq
+
+    def _cal_risk_analysis_scaler(self, freq):


We can combine these into risk_analysis and make it more powerful

you-n-g · 2021-05-08T05:46:51Z

qlib/utils/__init__.py

+            raise ValueError("sample freq must be xmin, xd, xw, xm")
+
+
+def get_sample_freq_calendar(start_time=None, end_time=None, freq="day", **kwargs):


Give some docs about the **kwargs.
What could it be?

you-n-g · 2021-05-08T06:24:50Z

qlib/utils/__init__.py

+    try:
+        _calendar = Cal.calendar(start_time=start_time, end_time=end_time, freq=freq, **kwargs)
+        freq, freq_sam = freq, None
+    except ValueError:


This part looks not intuitive.
Let's have some discussions later

qlib/contrib/strategy/rule_strategy.py

you-n-g · 2021-05-08T06:32:01Z

qlib/contrib/strategy/rule_strategy.py

+            self.instruments = D.instruments(instruments)
+        self.freq = freq
+
+    def _convert_index_format(self, df):


This function appears multiple times.
It will be better move it into utils.

you-n-g · 2021-05-08T06:55:28Z

qlib/contrib/backtest/account.py

+
+        for k, v in kwargs.items():
+            if hasattr(self, k):
+                setattr(self, k, v)


Give warning in else branch

bxdd · 2021-05-13T17:57:41Z

I add a notebook file workflow.ipynb to show the multi-level reports, its size is 5k. I temporarily modified .gitignore to submit it successfully. And this file has the same purpose as workflow_by_code.ipynb. Is this file necessary? If not, I will reset the related commit. @you-n-g

ultmaster · 2021-05-18T03:22:56Z

examples/multi_level_trading/workflow.py

+                "n_drop": 5,
+            },
+        },
+        "env": {


"env" -> "executor"?

ultmaster · 2021-05-18T03:30:35Z

examples/multi_level_trading/workflow.py

+                    "class": "SimulatorExecutor",
+                    "module_path": "qlib.contrib.backtest.executor",
+                    "kwargs": {
+                        "step_bar": "day",


Let's figure out a better name for step_bar and freq

ultmaster · 2021-05-18T03:49:33Z

qlib/contrib/backtest/executor.py

+        self._init_sub_trading(order_list)
+        sub_execute_state = self.sub_env.get_init_state()
+        while not self.sub_env.finished():
+            _order_list = self.sub_strategy.generate_order_list(sub_execute_state)


Why current can't be global?

Why can't we pass an account instead of sub_execute_state? account should contain everything that strategy needs other than information in exchange.

ultmaster · 2021-05-18T04:11:20Z

qlib/strategy/base.py

+from ..rl.interpreter import ActionInterpreter, StateInterpreter
+
+
+class BaseStrategy(BaseTradeCalendar):


Strategy should not stateless.

After reading the implementation of TWAP, I agree that the strategy is hard to be stateless.

ultmaster · 2021-05-18T04:14:10Z

qlib/strategy/base.py

+        _interpret_state = self.state_interpretor.interpret(
+            execute_result=execute_state, **self.action_interpret_kwargs
+        )
+        _policy_action = self.policy.step(_interpret_state)


self.policy(_interpret_state)

you-n-g · 2021-05-16T14:51:30Z

qlib/contrib/backtest/faculty.py

+# Licensed under the MIT License.
+
+
+class Faculty:


Why use this name?

you-n-g · 2021-05-17T12:27:56Z

qlib/contrib/backtest/faculty.py

+        self.__dict__["_faculty"].update(*args, **kwargs)
+
+
+common_faculty = Faculty()


Singleton is not enough for our scenario.

you-n-g · 2021-05-17T12:32:59Z

qlib/contrib/backtest/order.py

@@ -24,6 +24,7 @@ def __init__(self, stock_id, amount, trade_date, direction, factor):
        self.amount = amount
        # amount of successfully completed orders
        self.deal_amount = 0
-        self.trade_date = trade_date
+        self.start_time = start_time


define the start_time & end_time (e.g. include or exclude)

you-n-g · 2021-05-17T12:34:12Z

qlib/contrib/backtest/position.py


    def update_stock_price(self, stock_id, price):
        self.position[stock_id]["price"] = price

-    def update_stock_count(self, stock_id, count):
-        self.position[stock_id]["count"] = count
+    def update_stock_count(self, stock_id, bar, count):


will it better to unify the name bar and freq?

you-n-g · 2021-05-18T01:23:46Z

qlib/contrib/backtest/position.py

        del p["cash"]
-        del p["today_account_value"]
+        del p["now_account_value"]
        positions = pd.DataFrame.from_dict(p, orient="index")


Please import pandas and numpy

you-n-g · 2021-05-19T06:38:47Z

qlib/utils/sample.py

+        else:
+            if raw_count > sam_count:
+                raise ValueError("raw freq must be higher than sampling freq")
+        _calendar_minute = np.unique(


Why do we have to implement such a complicated version?
Will the following logic simpler?

div = freq_targert / freq_orig cal_target = cal_orig[::div]

add docstring

you-n-g · 2021-05-19T06:47:53Z

qlib/utils/sample.py

+        start sampling time, by default None
+    end_time : Union[str, pd.Timestamp], optional
+        end sampling time, by default None
+    fields : Union[str, List[str]], optional


What scenario do we have to resample part of the field?

you-n-g · 2021-05-19T06:49:22Z

qlib/utils/sample.py

+            else feature.loc[(slice(None), selector_datetime), fields]
+        )
+    if feature.empty:
+        return None


Returning the empty feature will be more reasonable

you-n-g · 2021-05-19T06:51:41Z

qlib/utils/sample.py

+    from ..data.dataset.utils import get_level_index
+
+    datetime_level = get_level_index(feature, level="datetime") == 0
+    if isinstance(feature, pd.Series):


If we don't filter fields in this function.
It will be unnecessary to use different logic between pd.Series and pd.DataFrame

you-n-g · 2021-05-19T06:53:11Z

qlib/utils/sample.py

+
+    from ..data.dataset.utils import get_level_index
+
+    datetime_level = get_level_index(feature, level="datetime") == 0


How do you make sure the datetime is sorted?

lasy_sort_index
qlib.utils

index.is_lexsorted()

you-n-g · 2021-05-19T07:28:22Z

examples/multi_level_trading/workflow.py

+                    "class": "SimulatorExecutor",
+                    "module_path": "qlib.contrib.backtest.executor",
+                    "kwargs": {
+                        "step_bar": "day",


Please send all the new names to the group for discussion

you-n-g · 2021-05-19T08:01:44Z

qlib/contrib/backtest/account.py

        self.current = Position(cash=init_cash)
+        self._reset_report()
+
+    def _cal_benchmark(self, benchmark_config, freq):


Move it to report.py

you-n-g · 2021-05-19T08:13:16Z

qlib/contrib/backtest/account.py

@@ -83,9 +165,13 @@ def update_order(self, order, trade_val, cost, trade_price):
            self.current.update_order(order, trade_val, cost, trade_price)
            self.update_state_from_order(order, trade_val, cost, trade_price)

-    def update_daily_end(self, today, trader):
+    def update_bar_count(self):


Giving Account an interface is doable

you-n-g · 2021-05-19T08:21:39Z

qlib/contrib/backtest/backtest.py


+    _execute_state = trade_env.get_init_state()
+    while not trade_env.finished():
+        _order_list = trade_strategy.generate_order_list(_execute_state)


for example, decision

you-n-g · 2021-05-19T08:25:21Z

qlib/contrib/backtest/backtest.py


+    _execute_state = trade_env.get_init_state()
+    while not trade_env.finished():
+        _order_list = trade_strategy.generate_order_list(_execute_state)


list sharing granularity and send it to group for discussion

you-n-g · 2021-05-19T08:28:35Z

qlib/contrib/backtest/exchange.py

@@ -51,6 +56,9 @@ def __init__(
                                                target on this day).
                                    index: MultipleIndex(instrument, pd.Datetime)
        """
+        self.freq = freq
+        self.start_time = start_time
+        self.end_time = end_time


1 2 3 4 5

1 2 3 4
[1, 4]
[1, 4.5]

you-n-g · 2021-05-19T09:21:11Z

qlib/contrib/backtest/exchange.py



 class Exchange:
    def __init__(
        self,
-        trade_dates=None,
+        freq="day",


turnover limit threshing

you-n-g · 2021-05-19T09:51:09Z

qlib/utils/sample.py

+        else:
+            if raw_count > sam_count:
+                raise ValueError("raw freq must be higher than sampling freq")
+        _calendar_minute = np.unique(


add docstring

you-n-g · 2021-05-19T09:56:07Z

qlib/utils/sample.py

+
+    from ..data.dataset.utils import get_level_index
+
+    datetime_level = get_level_index(feature, level="datetime") == 0


lasy_sort_index
qlib.utils

index.is_lexsorted()

…dd-qlib_highfreq_backtest

ultmaster · 2021-05-25T03:12:29Z

qlib/contrib/strategy/model_strategy.py

+        Return the proportion of your total value you will used in investment.
+        Dynamically risk_degree will result in Market timing.
+        """
+        # It will use 95% amoutn of your total value by default


ultmaster · 2021-05-25T03:14:01Z

qlib/strategy/base.py

+        if rely_trade_decision is not None:
+            self.rely_trade_decision = rely_trade_decision
+
+    def generate_trade_decision(self, execute_state):


What is the format/interface of execute state? What is expected to get from execute_state when I write a new strategy.

What is the interface of return value of generate_trade_decision?

ultmaster · 2021-05-25T03:19:49Z

qlib/strategy/base.py

+        if "trade_account" in common_infra:
+            self.trade_position = common_infra.get("trade_account").current
+
+    def reset(self, level_infra: dict = None, common_infra: dict = None, rely_trade_decision=None, **kwargs):


What is the interface of:

level_infra

common_infra

rely_trade_decision

When is reset expected to be called?

bxdd · 2021-05-25T07:08:27Z

This pr is closed, see #438

bxdd · 2021-05-27T12:50:23Z

qlib/utils/sample.py

@@ -35,7 +35,7 @@ def parse_freq(freq: str) -> Tuple[int, str]:
        raise ValueError(
            "freq format is not supported, the freq should be like (n)month/mon, (n)week/w, (n)day/d, (n)minute/min"
        )
-    _count = int(match_obj.group(1) if match_obj.group(1) else "1")
+    _count = int(match_obj.group(1)) if match_obj.group(1) is None else 1


Should it be int(match_obj.group(1)) if match_obj.group(1) else 1?
Example: If call parse_freq("min"), match_obj.group(1) == ' ' rather than None

bxdd added 17 commits March 20, 2021 00:11

add sample & base class

d3a1e03

update strategy

971d6a2

update report & account

8979d78

update env & strategy, add workflow

39deb7d

update trade calendar & backtest workflow

b14efa1

fix bug

af0053e

del outdate file

8920c19

trade_account support multi bar report

86a6f56

update port_ana_record

49cdaf8

black format

f404a03

fix bug in recorder

a109df3

fix bugs

d297a49

solve the conflict

e30df11

del old strategy

ae33950

fix trade time bug

7540ecd

black format

bc3eada

update the internal bar strategy

f7d3096

you-n-g reviewed May 8, 2021

View reviewed changes

Derek-Wds added the enhancement New feature or request label May 10, 2021

bxdd added 3 commits May 12, 2021 02:17

fix some comments and add docstring

621cb24

fix comments

07eaada

fix rule_strategy reset method

c703dab

bxdd requested a review from you-n-g May 12, 2021 16:50

bxdd added 2 commits May 13, 2021 22:39

fix rule_strategy bug

de2658a

update rule_startegy & add README, notebook for multi-level trading

ea60e60

optimize rule_strategy performance

eaa719d

ultmaster reviewed May 18, 2021

View reviewed changes

Update record_temp.py

dda509d

you-n-g reviewed May 19, 2021

View reviewed changes

Update sample.py

26d75b7

you-n-g reviewed May 19, 2021

View reviewed changes

you-n-g mentioned this pull request May 22, 2021

Enter a position on open price and exit on close price #432

Closed

bxdd added 3 commits May 25, 2021 02:38

fix comments

0c6e505

Merge branch 'qlib_highfreq_backtest' of github.com:bxdd/qlib into bx…

75fcb38

…dd-qlib_highfreq_backtest

solve the conflict

ee74489

bxdd requested review from rk2900, ultmaster and you-n-g May 24, 2021 18:57

ultmaster reviewed May 25, 2021

View reviewed changes

bxdd closed this May 25, 2021

bxdd commented May 27, 2021

View reviewed changes

		@@ -2,12 +2,12 @@
		# Licensed under the MIT License.

		from .order import Order

		raise ValueError("sample freq must be xmin, xd, xw, xm")


		def get_sample_freq_calendar(start_time=None, end_time=None, freq="day", **kwargs):

		from ..rl.interpreter import ActionInterpreter, StateInterpreter


		class BaseStrategy(BaseTradeCalendar):

		self.__dict__["_faculty"].update(args, *kwargs)


		common_faculty = Faculty()


		from ..data.dataset.utils import get_level_index

		datetime_level = get_level_index(feature, level="datetime") == 0

Support Highfreq Backtest with the Model/Rule/RL Strategy #408

Support Highfreq Backtest with the Model/Rule/RL Strategy #408

Conversation

bxdd commented Apr 30, 2021

Description

Motivation and Context

How Has This Been Tested?

Screenshots of Test Results (if appropriate):

Types of changes

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bxdd commented May 13, 2021

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

bxdd commented May 25, 2021

bxdd May 27, 2021 • edited Loading

Choose a reason for hiding this comment

bxdd May 27, 2021 •

edited

Loading