hroff-1902 6a397f579e

2019-09-29 00:43:27 +03:00

22 KiB

Raw Blame History

Hyperopt

This page explains how to tune your strategy by finding the optimal parameters, a process called hyperparameter optimization. The bot uses several algorithms included in the scikit-optimize package to accomplish this. The search will burn all your CPU cores, make your laptop sound like a fighter jet and still take a long time.

Hyperopt requires historic data to be available, just as backtesting does. To learn how to get data for the pairs and exchange you're interrested in, head over to the Data Downloading section of the documentation.

!!! Bug Hyperopt will crash when used with only 1 CPU Core as found out in Issue #1133

Prepare Hyperopting

Before we start digging into Hyperopt, we recommend you to take a look at an example hyperopt file located into user_data/hyperopts/

Configuring hyperopt is similar to writing your own strategy, and many tasks will be similar and a lot of code can be copied across from the strategy.

Checklist on all tasks / possibilities in hyperopt

Depending on the space you want to optimize, only some of the below are required:

fill populate_indicators - probably a copy from your strategy
fill buy_strategy_generator - for buy signal optimization
fill indicator_space - for buy signal optimzation
fill sell_strategy_generator - for sell signal optimization
fill sell_indicator_space - for sell signal optimzation

Optional, but recommended:

copy populate_buy_trend from your strategy - otherwise default-strategy will be used
copy populate_sell_trend from your strategy - otherwise default-strategy will be used

Rarely you may also need to override:

roi_space - for custom ROI optimization (if you need the ranges for the ROI parameters in the optimization hyperspace that differ from default)
generate_roi_table - for custom ROI optimization (if you need more than 4 entries in the ROI table)
stoploss_space - for custom stoploss optimization (if you need the range for the stoploss parameter in the optimization hyperspace that differs from default)

1. Install a Custom Hyperopt File

Put your hyperopt file into the directory user_data/hyperopts.

Let assume you want a hyperopt file awesome_hyperopt.py:
Copy the file user_data/hyperopts/sample_hyperopt.py into user_data/hyperopts/awesome_hyperopt.py

2. Configure your Guards and Triggers

There are two places you need to change in your hyperopt file to add a new buy hyperopt for testing:

Inside indicator_space() - the parameters hyperopt shall be optimizing.
Inside populate_buy_trend() - applying the parameters.

There you have two different types of indicators: 1. guards and 2. triggers.

Guards are conditions like "never buy if ADX < 10", or never buy if current price is over EMA10.
Triggers are ones that actually trigger buy in specific moment, like "buy when EMA5 crosses over EMA10" or "buy when close price touches lower bollinger band".

Hyperoptimization will, for each eval 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".

If you have updated the buy strategy, ie. changed the contents of populate_buy_trend() method you have to update the guards and triggers hyperopts must use.

Sell optimization

Similar to the buy-signal above, sell-signals can also be optimized. Place the corresponding settings into the following methods

Inside sell_indicator_space() - the parameters hyperopt shall be optimizing.
Inside populate_sell_trend() - applying the parameters.

The configuration and rules are the same than for buy signals. To avoid naming collisions in the search-space, please prefix all sell-spaces with sell-.

Using ticker-interval as part of the Strategy

The Strategy exposes the ticker-interval as self.ticker_interval. The same value is available as class-attribute HyperoptName.ticker_interval. In the case of the linked sample-value this would be SampleHyperOpts.ticker_interval.

Solving a Mystery

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 a search space:

    def indicator_space() -> List[Dimension]:
        """
        Define your Hyperopt space for searching strategy parameters
        """
        return [
            Integer(20, 40, name='adx-value'),
            Integer(20, 40, name='rsi-value'),
            Categorical([True, False], name='adx-enabled'),
            Categorical([True, False], name='rsi-enabled'),
            Categorical(['bb_lower', 'macd_cross_signal'], name='trigger')
        ]

Above definition says: I have five parameters I want you to randomly combine to find the best combination. Two of them are integer values (adx-value and rsi-value) and I want you test in the range of values 20 to 40. Then we have three category variables. First two are either True or False. We use these to either enable or disable the ADX and RSI guards. The last one we call trigger and use it to decide which buy trigger we want to use.

So let's write the buy strategy using these values:

        def populate_buy_trend(dataframe: DataFrame) -> DataFrame:
            conditions = []
            # GUARDS AND TRENDS
            if 'adx-enabled' in params and params['adx-enabled']:
                conditions.append(dataframe['adx'] > params['adx-value'])
            if 'rsi-enabled' in params and params['rsi-enabled']:
                conditions.append(dataframe['rsi'] < params['rsi-value'])

            # TRIGGERS
            if 'trigger' in params:
                if params['trigger'] == 'bb_lower':
                    conditions.append(dataframe['close'] < dataframe['bb_lowerband'])
                if params['trigger'] == 'macd_cross_signal':
                    conditions.append(qtpylib.crossed_above(
                        dataframe['macd'], dataframe['macdsignal']
                    ))

            if conditions:
                dataframe.loc[
                    reduce(lambda x, y: x & y, conditions),
                    'buy'] = 1

            return dataframe

        return populate_buy_trend

Hyperopting will now call this populate_buy_trend as many times you ask it (epochs) with different value combinations. It will then use the given historical data and make buys based on the buy signals generated with the above function and based on the results it will end with telling you which paramter combination produced the best profits.

The search for best parameters starts with a few random combinations and then uses a regressor algorithm (currently ExtraTreesRegressor) to quickly find a parameter combination that minimizes the value of the loss function.

The above setup expects to find ADX, RSI and Bollinger Bands in the populated indicators. When you want to test an indicator that isn't used by the bot currently, remember to add it to the populate_indicators() method in hyperopt.py.

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.

By default, FreqTrade uses a loss function, which has been with freqtrade since the beginning and optimizes mostly for short trade duration and avoiding losses.

A different loss function can be specified by using the --hyperopt-loss <Class-name> argument. This class should be in its own file within the user_data/hyperopts/ directory.

Currently, the following loss functions are builtin:

DefaultHyperOptLoss (default legacy Freqtrade hyperoptimization loss function)
OnlyProfitHyperOptLoss (which takes only amount of profit into consideration)
SharpeHyperOptLoss (optimizes Sharpe Ratio calculated on the trade returns)

Creating and using a custom loss function

To use a custom loss function class, make sure that the function hyperopt_loss_function is defined in your custom hyperopt loss class. For the sample below, you then need to add the command line parameter --hyperopt-loss SuperDuperHyperOptLoss to your hyperopt call so this fuction is being used.

A sample of this can be found below, which is identical to the Default Hyperopt loss implementation. A full sample can be found user_data/hyperopts/

from freqtrade.optimize.hyperopt import IHyperOptLoss

TARGET_TRADES = 600
EXPECTED_MAX_PROFIT = 3.0
MAX_ACCEPTED_TRADE_DURATION = 300

class SuperDuperHyperOptLoss(IHyperOptLoss):
    """
    Defines the default loss function for hyperopt
    """

    @staticmethod
    def hyperopt_loss_function(results: DataFrame, trade_count: int,
                               min_date: datetime, max_date: datetime,
                               *args, **kwargs) -> float:
        """
        Objective function, returns smaller number for better results
        This is the legacy algorithm (used until now in freqtrade).
        Weights are distributed as follows:
        * 0.4 to trade duration
        * 0.25: Avoiding trade loss
        * 1.0 to total profit, compared to the expected value (`EXPECTED_MAX_PROFIT`) defined above
        """
        total_profit = results.profit_percent.sum()
        trade_duration = results.trade_duration.mean()

        trade_loss = 1 - 0.25 * exp(-(trade_count - TARGET_TRADES) ** 2 / 10 ** 5.8)
        profit_loss = max(0, 1 - total_profit / EXPECTED_MAX_PROFIT)
        duration_loss = 0.4 * min(trade_duration / MAX_ACCEPTED_TRADE_DURATION, 1)
        result = trade_loss + profit_loss + duration_loss
        return result

Currently, the arguments are:

results: DataFrame containing the result
The following columns are available in results (corresponds to the output-file of backtesting when used with --export trades):
pair, profit_percent, profit_abs, open_time, close_time, open_index, close_index, trade_duration, open_at_end, open_rate, close_rate, sell_reason
trade_count: Amount of trades (identical to len(results))
min_date: Start date of the hyperopting TimeFrame
min_date: End date of the hyperopting TimeFrame

This function needs to return a floating point number (float). Smaller numbers will be interpreted as better results. The parameters and balancing for this is up to you.

!!! Note This function is called once per iteration - so please make sure to have this as optimized as possible to not slow hyperopt down unnecessarily.

!!! Note Please keep the arguments *args and **kwargs in the interface to allow us to extend this interface later.

Execute Hyperopt

Once you have updated your hyperopt configuration you can run it. Because hyperopt tries a lot of combinations to find the best parameters it will take time to get a good result. More time usually results in better results.

We strongly recommend to use screen or tmux to prevent any connection loss.

freqtrade -c config.json hyperopt --customhyperopt <hyperoptname> -e 5000 --spaces all

Use <hyperoptname> as the name of the custom hyperopt used.

The -e flag will set how many evaluations hyperopt will do. We recommend running at least several thousand evaluations.

The --spaces all flag determines that all possible parameters should be optimized. Possibilities are listed below.

!!! Note By default, hyperopt will erase previous results and start from scratch. Continuation can be archived by using --continue.

!!! Warning When switching parameters or changing configuration options, make sure to not use the argument --continue so temporary results can be removed.

Execute Hyperopt with Different Ticker-Data Source

If you would like to hyperopt parameters using an alternate ticker data that you have on-disk, use the --datadir PATH option. Default hyperopt will use data from directory user_data/data.

Running Hyperopt with Smaller Testset

Use the --timerange argument to change how much of the testset you want to use. For example, to use one month of data, pass the following parameter to the hyperopt call:

freqtrade hyperopt --timerange 20180401-20180501

Running Hyperopt with Smaller Search Space

Use the --spaces argument to limit the search space used by hyperopt. Letting Hyperopt optimize everything is a huuuuge search space. Often it might make more sense to start by just searching for initial buy algorithm. Or maybe you just want to optimize your stoploss or roi table for that awesome new buy strategy you have.

Legal values are:

all: optimize everything
buy: just search for a new buy strategy
sell: just search for a new sell strategy
roi: just optimize the minimal profit table for your strategy
stoploss: search for the best stoploss value
space-separated list of any of the above values for example --spaces roi stoploss

Position stacking and disabling max market positions

In some situations, you may need to run Hyperopt (and Backtesting) with the --eps/--enable-position-staking and --dmmp/--disable-max-market-positions arguments.

By default, hyperopt emulates the behavior of the Freqtrade Live Run/Dry Run, where only one open trade is allowed for every traded pair. The total number of trades open for all pairs is also limited by the max_open_trades setting. During Hyperopt/Backtesting this may lead to some potential trades to be hidden (or masked) by previosly open trades.

The --eps/--enable-position-stacking argument allows emulation of buying the same pair multiple times, while --dmmp/--disable-max-market-positions disables applying max_open_trades during Hyperopt/Backtesting (which is equal to setting max_open_trades to a very high number).

!!! Note Dry/live runs will NOT use position stacking - therefore it does make sense to also validate the strategy without this as it's closer to reality.

You can also enable position stacking in the configuration file by explicitly setting "position_stacking"=true.

Understand the Hyperopt Result

Once Hyperopt is completed you can use the result to create a new strategy. Given the following result from hyperopt:

Best result:

    44/100:    135 trades. Avg profit  0.57%. Total profit  0.03871918 BTC (0.7722Σ%). Avg duration 180.4 mins. Objective: 1.94367

Buy hyperspace params:
{    'adx-value': 44,
     'rsi-value': 29,
     'adx-enabled': False,
     'rsi-enabled': True,
     'trigger': 'bb_lower'}

You should understand this result like:

The buy trigger that worked best was bb_lower.
You should not use ADX because adx-enabled: False)
You should consider using the RSI indicator (rsi-enabled: True and the best value is 29.0 (rsi-value: 29.0)

You have to look inside your strategy file into buy_strategy_generator() method, what those values match to.

So for example you had rsi-value: 29.0 so we would look at rsi-block, that translates to the following code block:

(dataframe['rsi'] < 29.0)

Translating your whole hyperopt result as the new buy-signal would then look like:

def populate_buy_trend(self, dataframe: DataFrame) -> DataFrame:
    dataframe.loc[
        (
            (dataframe['rsi'] < 29.0) &  # rsi-value
            dataframe['close'] < dataframe['bb_lowerband']  # trigger
        ),
        'buy'] = 1
    return dataframe

By default, hyperopt prints colorized results -- epochs with positive profit are printed in the green color. This highlighting helps you find epochs that can be interesting for later analysis. Epochs with zero total profit or with negative profits (losses) are printed in the normal color. If you do not need colorization of results (for instance, when you are redirecting hyperopt output to a file) you can switch colorization off by specifying the --no-color option in the command line.

You can use the --print-all command line option if you would like to see all results in the hyperopt output, not only the best ones. When --print-all is used, current best results are also colorized by default -- they are printed in bold (bright) style. This can also be switched off with the --no-color command line option.

Understand Hyperopt ROI results

If you are optimizing ROI (i.e. if optimization search-space contains 'all' or 'roi'), your result will look as follows and include a ROI table:

Best result:

    44/100:    135 trades. Avg profit  0.57%. Total profit  0.03871918 BTC (0.7722Σ%). Avg duration 180.4 mins. Objective: 1.94367

Buy hyperspace params:
{   'adx-value': 44,
    'rsi-value': 29,
    'adx-enabled': False,
    'rsi-enabled': True,
    'trigger': 'bb_lower'}
ROI table:
{   0: 0.10674,
    21: 0.09158,
    78: 0.03634,
    118: 0}

In order to use this best ROI table found by Hyperopt in backtesting and for live trades/dry-run, copy-paste it as the value of the minimal_roi attribute of your custom strategy:

    # Minimal ROI designed for the strategy.
    # This attribute will be overridden if the config file contains "minimal_roi"
    minimal_roi = {
        0: 0.10674,
        21: 0.09158,
        78: 0.03634,
        118: 0
    }

As stated in the comment, you can also use it as the value of the minimal_roi setting in the configuration file.

Default ROI Search Space

If you are optimizing ROI, Freqtrade creates the 'roi' optimization hyperspace for you -- it's the hyperspace of components for the ROI tables. By default, each ROI table generated by the Freqtrade consists of 4 rows (steps). Hyperopt implements adaptive ranges for ROI tables with ranges for values in the ROI steps that depend on the ticker_interval used. By default the values can vary in the following ranges (for some of the most used ticker intervals, values are rounded to 5 digits after the decimal point):

# step	1m		5m		1h		1d
1	0	0.01161...0.11992	0	0.03...0.31	0	0.06883...0.71124	0	0.12178...1.25835
2	2...8	0.00774...0.04255	10...40	0.02...0.11	120...480	0.04589...0.25238	2880...11520	0.08118...0.44651
3	4...20	0.00387...0.01547	20...100	0.01...0.04	240...1200	0.02294...0.09177	5760...28800	0.04059...0.16237
4	6...44	0.0	30...220	0.0	360...2640	0.0	8640...63360	0.0

These ranges should be sufficient in most cases. The minutes in the steps (ROI dict keys) are scaled linearly depending on the ticker interval used. The ROI values in the steps (ROI dict values) are scaled logarithmically depending on the ticker interval used.

If you have the generate_roi_table() and roi_space() methods in your custom hyperopt file, remove them in order to utilize these adaptive ROI tables and the ROI hyperoptimization space generated by Freqtrade by default.

Override the roi_space() method if you need components of the ROI tables to vary in other ranges. Override the generate_roi_table() and roi_space() methods and implement your own custom approach for generation of the ROI tables during hyperoptimization if you need a different structure of the ROI tables or other amount of rows (steps). A sample for these methods can be found in user_data/hyperopts/sample_hyperopt_advanced.py.

Understand Hyperopt Stoploss results

If you are optimizing stoploss values (i.e. if optimization search-space contains 'all' or 'stoploss'), your result will look as follows and include stoploss:

Best result:

    44/100:    135 trades. Avg profit  0.57%. Total profit  0.03871918 BTC (0.7722Σ%). Avg duration 180.4 mins. Objective: 1.94367

Buy hyperspace params:
{   'adx-value': 44,
    'rsi-value': 29,
    'adx-enabled': False,
    'rsi-enabled': True,
    'trigger': 'bb_lower'}
Stoploss: -0.27996

In order to use this best stoploss value found by Hyperopt in backtesting and for live trades/dry-run, copy-paste it as the value of the stoploss attribute of your custom strategy:

    # Optimal stoploss designed for the strategy
    # This attribute will be overridden if the config file contains "stoploss"
    stoploss = -0.27996

As stated in the comment, you can also use it as the value of the stoploss setting in the configuration file.

Default Stoploss Search Space

If you are optimizing stoploss values, Freqtrade creates the 'stoploss' optimization hyperspace for you. By default, the stoploss values in that hyperspace can vary in the range -0.35...-0.02, which is sufficient in most cases.

If you have the stoploss_space() method in your custom hyperopt file, remove it in order to utilize Stoploss hyperoptimization space generated by Freqtrade by default.

Override the stoploss_space() method and define the desired range in it if you need stoploss values to vary in other range during hyperoptimization. A sample for this method can be found in user_data/hyperopts/sample_hyperopt_advanced.py.

Validate backtesting results

Once the optimized strategy has been implemented into your strategy, you should backtest this strategy to make sure everything is working as expected.

To achieve same results (number of trades, their durations, profit, etc.) than during Hyperopt, please use same set of arguments --dmmp/--disable-max-market-positions and --eps/--enable-position-stacking for Backtesting.

Next Step

Now you have a perfect bot and want to control it from Telegram. Your next step is to learn the Telegram usage.

22 KiB Raw Blame History