adelorenzo/stable
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #4790 from freqtrade/hyper_range

Browse Source 
add IntParameter.range
This commit is contained in:
Matthias 2021-04-24 13:41:29 +02:00 committed by GitHub
commit 9177d9d129
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 192 additions and 48 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

181
docs/hyperopt.md
View File

@ -165,11 +165,22 @@ Rarely you may also need to create a [nested class](advanced-hyperopt.md#overrid
!!! Tip "Quickly optimize ROI, stoploss and trailing stoploss"
You can quickly optimize the spaces `roi`, `stoploss` and `trailing` without changing anything in your strategy.
```python
``` bash
# Have a working strategy at hand.
freqtrade hyperopt --hyperopt-loss SharpeHyperOptLossDaily --spaces roi stoploss trailing --strategy MyWorkingStrategy --config config.json -e 100
```
### Hyperopt execution logic
Hyperopt will first load your data into memory and will then run `populate_indicators()` once per Pair to generate all indicators.
Hyperopt will then spawn into different processes (number of processors, or `-j <n>`), and run backtesting over and over again, changing the parameters that are part of the `--spaces` defined.
For every new set of parameters, freqtrade will run first `populate_buy_trend()` followed by `populate_sell_trend()`, and then run the regular backtesting process to simulate trades.
After backtesting, the results are passed into the [loss function](#loss-functions), which will evaluate if this result was better or worse than previous results.
Based on the loss function result, hyperopt will determine the next set of parameters to try in the next round of backtesting.
### Configure your Guards and Triggers
There are two places you need to change in your strategy file to add a new buy hyperopt for testing:
@ -187,60 +198,54 @@ There you have two different types of indicators: 1. `guards` and 2. `triggers`.
However, this guide will make this distinction to make it clear that signals should not be "sticking".
Sticking signals are signals that are active for multiple candles. This can lead into buying a signal late (right before the signal disappears - which means that the chance of success is a lot lower than right at the beginning).
Hyper-optimization will, for each epoch round, pick one trigger and possibly
multiple guards. The constructed strategy will be something like "*buy exactly when close price touches lower Bollinger band, BUT only if
ADX > 10*".
```python
from freqtrade.strategy import IntParameter, IStrategy
class MyAwesomeStrategy(IStrategy):
# If parameter is prefixed with `buy_` or `sell_` then specifying `space` parameter is optional
# and space is inferred from parameter name.
buy_adx_min = IntParameter(0, 100, default=10)
def populate_buy_trend(self, dataframe: 'DataFrame', metadata: dict) -> 'DataFrame':
dataframe.loc[
(
(dataframe['adx'] > self.buy_adx_min.value)
), 'buy'] = 1
return dataframe
```
Hyper-optimization will, for each epoch round, pick one trigger and possibly multiple guards.
#### Sell optimization
Similar to the buy-signal above, sell-signals can also be optimized.
Place the corresponding settings into the following methods
* Define the parameters at the class level hyperopt shall be optimizing.
* Define the parameters at the class level hyperopt shall be optimizing, either naming them `sell_*`, or by explicitly defining `space='sell'`.
* Within `populate_sell_trend()` - use defined parameter values instead of raw constants.
The configuration and rules are the same than for buy signals.
```python
class MyAwesomeStrategy(IStrategy):
# There is no strict parameter naming scheme. If you do not use `buy_` or `sell_` prefixes -
# please specify to which space parameter belongs using `space` parameter. Possible values:
# 'buy' or 'sell'.
adx_max = IntParameter(0, 100, default=50, space='sell')
## Solving a Mystery
def populate_sell_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
dataframe.loc[
(
(dataframe['adx'] < self.adx_max.value)
), 'buy'] = 1
Let's say you are curious: should you use MACD crossings or lower Bollinger Bands to trigger your buys.
And you also wonder should you use RSI or ADX to help with those buy decisions.
If you decide to use RSI or ADX, which values should I use for them?
So let's use hyperparameter optimization to solve this mystery.
### Defining indicators to be used
We start by calculating the indicators our strategy is going to use.
``` python
class MyAwesomeStrategy(IStrategy):
def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
"""
Generate all indicators used by the strategy
"""
dataframe['adx'] = ta.ADX(dataframe)
dataframe['rsi'] = ta.RSI(dataframe)
macd = ta.MACD(dataframe)
dataframe['macd'] = macd['macd']
dataframe['macdsignal'] = macd['macdsignal']
dataframe['macdhist'] = macd['macdhist']
bollinger = ta.BBANDS(dataframe, timeperiod=20, nbdevup=2.0, nbdevdn=2.0)
dataframe['bb_lowerband'] = boll['lowerband']
dataframe['bb_middleband'] = boll['middleband']
dataframe['bb_upperband'] = boll['upperband']
return dataframe
```
## Solving a Mystery
### Hyperoptable parameters
Let's say you are curious: should you use MACD crossings or lower Bollinger
Bands to trigger your buys. And you also wonder should you use RSI or ADX to
help with those buy decisions. If you decide to use RSI or ADX, which values
should I use for them? So let's use hyperparameter optimization to solve this
mystery.
We will start by defining hyperoptable parameters:
We continue to define hyperoptable parameters:
```python
class MyAwesomeStrategy(IStrategy):
@ -288,7 +293,7 @@ So let's write the buy strategy using these values:
```
Hyperopt will now call `populate_buy_trend()` many times (`epochs`) with different value combinations.
It will use the given historical data and make buys based on the buy signals generated with the above function.
It will use the given historical data and simulate buys based on the buy signals generated with the above function.
Based on the results, hyperopt will tell you which parameter combination produced the best results (based on the configured [loss function](#loss-functions)).
!!! Note
@ -314,6 +319,90 @@ There are four parameter types each suited for different purposes.
!!! Warning
Hyperoptable parameters cannot be used in `populate_indicators` - as hyperopt does not recalculate indicators for each epoch, so the starting value would be used in this case.
### Optimizing an indicator parameter
Assuming you have a simple strategy in mind - a EMA cross strategy (2 Moving averages crossing) - and you'd like to find the ideal parameters for this strategy.
``` python
from pandas import DataFrame
from functools import reduce
import talib.abstract as ta
from freqtrade.strategy import IStrategy
from freqtrade.strategy import CategoricalParameter, DecimalParameter, IntParameter
import freqtrade.vendor.qtpylib.indicators as qtpylib
class MyAwesomeStrategy(IStrategy):
stoploss = -0.05
timeframe = '15m'
# Define the parameter spaces
buy_ema_short = IntParameter(3, 50, default=5)
buy_ema_long = IntParameter(15, 200, default=50)
def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
"""Generate all indicators used by the strategy"""
# Calculate all ema_short values
for val in self.buy_ema_short.range:
dataframe[f'ema_short_{val}'] = ta.EMA(dataframe, timeperiod=val)
# Calculate all ema_long values
for val in self.buy_ema_long.range:
dataframe[f'ema_long_{val}'] = ta.EMA(dataframe, timeperiod=val)
return dataframe
def populate_buy_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
conditions = []
conditions.append(qtpylib.crossed_above(
dataframe[f'ema_short_{self.buy_ema_short.value}'], dataframe[f'ema_long_{self.buy_ema_long.value}']
))
# Check that volume is not 0
conditions.append(dataframe['volume'] > 0)
if conditions:
dataframe.loc[
reduce(lambda x, y: x & y, conditions),
'buy'] = 1
return dataframe
def populate_sell_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
conditions = []
conditions.append(qtpylib.crossed_above(
dataframe[f'ema_long_{self.buy_ema_long.value}'], dataframe[f'ema_short_{self.buy_ema_short.value}']
))
# Check that volume is not 0
conditions.append(dataframe['volume'] > 0)
if conditions:
dataframe.loc[
reduce(lambda x, y: x & y, conditions),
'sell'] = 1
return dataframe
```
Breaking it down:
Using `self.buy_ema_short.range` will return a range object containing all entries between the Parameters low and high value.
In this case (`IntParameter(3, 50, default=5)`), the loop would run for all numbers between 3 and 50 (`[3, 4, 5, ... 49, 50]`).
By using this in a loop, hyperopt will generate 48 new columns (`['buy_ema_3', 'buy_ema_4', ... , 'buy_ema_50']`).
Hyperopt itself will then use the selected value to create the buy and sell signals
While this strategy is most likely too simple to provide consistent profit, it should serve as an example how optimize indicator parameters.
!!! Note
`self.buy_ema_short.range` will act differently between hyperopt and other modes. For hyperopt, the above example may generate 48 new columns, however for all other modes (backtesting, dry/live), it will only generate the column for the selected value. You should therefore avoid using the resulting column with explicit values (values other than `self.buy_ema_short.value`).
??? Hint "Performance tip"
By doing the calculation of all possible indicators in `populate_indicators()`, the calculation of the indicator happens only once for every parameter.
While this may slow down the hyperopt startup speed, the overall performance will increase as the Hyperopt execution itself may pick the same value for multiple epochs (changing other values).
You should however try to use space ranges as small as possible. Every new column will require more memory, and every possibility hyperopt can try will increase the search space.
## Loss-functions
Each hyperparameter tuning requires a target. This is usually defined as a loss function (sometimes also called objective function), which should decrease for more desirable results, and increase for bad results.
@ -606,6 +695,16 @@ number).
You can also enable position stacking in the configuration file by explicitly setting
`"position_stacking"=true`.
## Out of Memory errors
As hyperopt consumes a lot of memory (the complete data needs to be in memory once per parallel backtesting process), it's likely that you run into "out of memory" errors.
To combat these, you have multiple options:
* reduce the amount of pairs
* reduce the timerange used (`--timerange <timerange>`)
* reduce the number of parallel processes (`-j <n>`)
* Increase the memory of your machine
## Show details of Hyperopt results
After you run Hyperopt for the desired amount of epochs, you can later list all results for analysis, select only best or profitable once, and show the details for any of the epochs previously evaluated. This can be done with the `hyperopt-list` and `hyperopt-show` sub-commands. The usage of these sub-commands is described in the [Utils](utils.md#list-hyperopt-results) chapter.

40
freqtrade/strategy/hyper.py
View File

@ -5,7 +5,7 @@ This module defines a base class for auto-hyperoptable strategies.
import logging
from abc import ABC, abstractmethod
from contextlib import suppress
from typing import Any, Iterator, Optional, Sequence, Tuple, Union
from typing import Any, Dict, Iterator, Optional, Sequence, Tuple, Union
with suppress(ImportError):
@ -13,6 +13,7 @@ with suppress(ImportError):
from freqtrade.optimize.space import SKDecimal
from freqtrade.exceptions import OperationalException
from freqtrade.state import RunMode
logger = logging.getLogger(__name__)
@ -25,6 +26,7 @@ class BaseParameter(ABC):
category: Optional[str]
default: Any
value: Any
hyperopt: bool = False
def __init__(self, *, default: Any, space: Optional[str] = None,
optimize: bool = True, load: bool = True, **kwargs):
@ -121,6 +123,20 @@ class IntParameter(NumericParameter):
"""
return Integer(low=self.low, high=self.high, name=name, **self._space_params)
@property
def range(self):
"""
Get each value in this space as list.
Returns a List from low to high (inclusive) in Hyperopt mode.
Returns a List with 1 item (`value`) in "non-hyperopt" mode, to avoid
calculating 100ds of indicators.
"""
if self.hyperopt:
# Scikit-optimize ranges are "inclusive", while python's "range" is exclusive
return range(self.low, self.high + 1)
else:
return range(self.value, self.value + 1)
class RealParameter(NumericParameter):
default: float
@ -227,12 +243,11 @@ class HyperStrategyMixin(object):
strategy logic.
"""
def __init__(self, *args, **kwargs):
def __init__(self, config: Dict[str, Any], *args, **kwargs):
"""
Initialize hyperoptable strategy mixin.
"""
self._load_params(getattr(self, 'buy_params', None))
self._load_params(getattr(self, 'sell_params', None))
self._load_hyper_params(config.get('runmode') == RunMode.HYPEROPT)
def enumerate_parameters(self, category: str = None) -> Iterator[Tuple[str, BaseParameter]]:
"""
@ -254,18 +269,29 @@ class HyperStrategyMixin(object):
(attr_name.startswith(category + '_') and attr.category is None)):
yield attr_name, attr
def _load_params(self, params: dict) -> None:
def _load_hyper_params(self, hyperopt: bool = False) -> None:
"""
Load Hyperoptable parameters
"""
self._load_params(getattr(self, 'buy_params', None), 'buy', hyperopt)
self._load_params(getattr(self, 'sell_params', None), 'sell', hyperopt)
def _load_params(self, params: dict, space: str, hyperopt: bool = False) -> None:
"""
Set optimizeable parameter values.
:param params: Dictionary with new parameter values.
"""
if not params:
return
logger.info(f"No params for {space} found, using default values.")
for attr_name, attr in self.enumerate_parameters():
if attr_name in params:
attr.hyperopt = hyperopt
if params and attr_name in params:
if attr.load:
attr.value = params[attr_name]
logger.info(f'Strategy Parameter: {attr_name} = {attr.value}')
else:
logger.warning(f'Parameter "{attr_name}" exists, but is disabled. '
f'Default value "{attr.value}" used.')
else:
logger.info(f'Strategy Parameter(default): {attr_name} = {attr.value}')

2
tests/optimize/conftest.py
View File

@ -6,6 +6,7 @@ import pandas as pd
import pytest
from freqtrade.optimize.hyperopt import Hyperopt
from freqtrade.state import RunMode
from freqtrade.strategy.interface import SellType
from tests.conftest import patch_exchange
@ -15,6 +16,7 @@ def hyperopt_conf(default_conf):
hyperconf = deepcopy(default_conf)
hyperconf.update({
'datadir': Path(default_conf['datadir']),
'runmode': RunMode.HYPEROPT,
'hyperopt': 'DefaultHyperOpt',
'hyperopt_loss': 'ShortTradeDurHyperOptLoss',
'hyperopt_path': str(Path(__file__).parent / 'hyperopts'),

9
tests/optimize/test_hyperopt.py
View File

@ -21,6 +21,7 @@ from freqtrade.optimize.hyperopt_tools import HyperoptTools
from freqtrade.optimize.space import SKDecimal
from freqtrade.resolvers.hyperopt_resolver import HyperOptResolver
from freqtrade.state import RunMode
from freqtrade.strategy.hyper import IntParameter
from tests.conftest import (get_args, log_has, log_has_re, patch_exchange,
patched_configuration_load_config_file)
@ -1103,6 +1104,14 @@ def test_in_strategy_auto_hyperopt(mocker, hyperopt_conf, tmpdir) -> None:
})
hyperopt = Hyperopt(hyperopt_conf)
assert isinstance(hyperopt.custom_hyperopt, HyperOptAuto)
assert isinstance(hyperopt.backtesting.strategy.buy_rsi, IntParameter)
assert hyperopt.backtesting.strategy.buy_rsi.hyperopt is True
assert hyperopt.backtesting.strategy.buy_rsi.value == 35
buy_rsi_range = hyperopt.backtesting.strategy.buy_rsi.range
assert isinstance(buy_rsi_range, range)
# Range from 0 - 50 (inclusive)
assert len(list(buy_rsi_range)) == 51
hyperopt.start()

8
tests/strategy/test_interface.py
View File

@ -588,6 +588,14 @@ def test_hyperopt_parameters():
intpar = IntParameter(low=0, high=5, default=1, space='buy')
assert intpar.value == 1
assert isinstance(intpar.get_space(''), Integer)
assert isinstance(intpar.range, range)
assert len(list(intpar.range)) == 1
# Range contains ONLY the default / value.
assert list(intpar.range) == [intpar.value]
intpar.hyperopt = True
assert len(list(intpar.range)) == 6
assert list(intpar.range) == [0, 1, 2, 3, 4, 5]
fltpar = RealParameter(low=0.0, high=5.5, default=1.0, space='buy')
assert isinstance(fltpar.get_space(''), Real)