Risk and Performance Metrics

The risk and performance metrics are summarizing values calculated by Zipline when running a simulation. These metrics can be about the performance of an algorithm, like returns or cash flow, or the riskiness of an algorithm, like volatility or beta. Metrics may be reported minutely, daily, or once at the end of a simulation. A single metric may choose to report at multiple time-scales where appropriate.

Metrics Sets

Zipline groups risk and performance metrics into collections called “metrics sets”. A single metrics set defines all of the metrics to track during a single backtest. A metrics set may contain metrics that report at different time scales. The default metrics set will compute a host of metrics, such as algorithm returns, volatility, Sharpe ratio, and beta.

Selecting the Metrics Set

When running a simulation, the user may select the metrics set to report. How you select the metrics set depends on the interface being used to run the algorithm.

Command Line and IPython Magic

When running with the command line or IPython magic interfaces, the metrics set may be selected by passing the --metrics-set argument. For example:

$ zipline run algorithm.py -s 2014-01-01 -e 2014-02-01 --metrics-set my-metrics-set

run_algorithm

When running through the run_algorithm() interface, the metrics set may be passed with the metrics_set argument. This may either be the name of a registered metrics set, or a set of metric object. For example:

run_algorithm(..., metrics_set='my-metrics-set')
run_algorithm(..., metrics_set={MyMetric(), MyOtherMetric(), ...})

Running Without Metrics

Computing risk and performance metrics is not free, and contributes to the total runtime of a backtest. When actively developing an algorithm, it is often helpful to skip these computations to speed up the debugging cycle. To disable the calculation and reporting of all metrics, users may select the built-in metrics set none. For example:

$ zipline run algorithm.py -s 2014-01-01 -e 2014-02-01 --metrics-set none

Defining New Metrics

A metric is any object that implements some subset of the following methods:

  • start_of_simulation
  • end_of_simulation
  • start_of_session
  • end_of_session
  • end_of_bar

These functions will be called at the time indicated by their name, at which point the metric object may collect any needed information and optionally report a computed value. If a metric does not need to do any processing at one of these times, it may omit a definition for the given method.

A metric should be reusable, meaning that a single instance of a metric class should be able to be used across multiple backtests. Metrics do not need to support multiple simulations at once, meaning that internal caches and data are consistent between start_of_simulation and end_of_simulation.

start_of_simulation

The start_of_simulation method should be thought of as a per-simulation constructor. This method should initialize any caches needed for the duration of a single simulation.

The start_of_simulation method should have the following signature:

def start_of_simulation(self,
                        ledger,
                        emission_rate,
                        trading_calendar,
                        sessions,
                        benchmark_source):
    ...

ledger is an instance of Ledger which is maintaining the simulation’s state. This may be used to lookup the algorithm’s starting portfolio values.

emission_rate is a string representing the smallest frequency at which metrics should be reported. emission_rate will be either minute or daily. When emission_rate is daily, end_of_bar will not be called at all.

trading_calendar is an instance of TradingCalendar which is the trading calendar being used by the simulation.

sessions is a pandas.DatetimeIndex which is holds the session labels, in sorted order, that the simulation will execute.

benchmark_source is an instance of BenchmarkSource which is the interface to the returns of the benchmark specified by set_benchmark().

end_of_simulation

The end_of_simulation method should have the following signature:

def end_of_simulation(self,
                      packet,
                      ledger,
                      trading_calendar,
                      sessions,
                      data_portal,
                      benchmark_source):
    ...

ledger is an instance of Ledger which is maintaining the simulation’s state. This may be used to lookup the algorithm’s final portfolio values.

packet is a dictionary to write the end of simulation values for the given metric into.

trading_calendar is an instance of TradingCalendar which is the trading calendar being used by the simulation.

sessions is a pandas.DatetimeIndex which is holds the session labels, in sorted order, that the simulation has executed.

data_portal is an instance of DataPortal which is the metric’s interface to pricing data.

benchmark_source is an instance of BenchmarkSource which is the interface to the returns of the benchmark specified by set_benchmark().

start_of_session

The start_of_session method may see a slightly different view of the ledger or data_portal than the previous end_of_session if the price of any futures owned move between trading sessions or if a capital change occurs.

The start_of_session method should have the following signature:

def start_of_session(self,
                     ledger,
                     session_label,
                     data_portal):
    ...

ledger is an instance of Ledger which is maintaining the simulation’s state. This may be used to lookup the algorithm’s current portfolio values.

session_label is a Timestamp which is the label of the session which is about to run.

data_portal is an instance of DataPortal which is the metric’s interface to pricing data.

end_of_session

The end_of_session method should have the following signature:

def end_of_session(self,
                   packet,
                   ledger,
                   session_label,
                   session_ix,
                   data_portal):

packet is a dictionary to write the end of session values. This dictionary contains two sub-dictionaries: daily_perf and cumulative_perf. When applicable, the daily_perf should hold the current day’s value, and cumulative_perf should hold a cumulative value for the entire simulation up to the current time.

ledger is an instance of Ledger which is maintaining the simulation’s state. This may be used to lookup the algorithm’s current portfolio values.

session_label is a Timestamp which is the label of the session which is has just completed.

session_ix is an int which is the index of the current trading session being run. This is provided to allow for efficient access to the daily returns through ledger.daily_returns_array[:session_ix + 1].

data_portal is an instance of DataPortal which is the metric’s interface to pricing data

end_of_bar

Note

end_of_bar is only called when emission_mode is minute.

The end_of_bar method should have the following signature:

def end_of_bar(self,
               packet,
               ledger,
               dt,
               session_ix,
               data_portal):

packet is a dictionary to write the end of session values. This dictionary contains two sub-dictionaries: minute_perf and cumulative_perf. When applicable, the minute_perf should hold the current partial day’s value, and cumulative_perf should hold a cumulative value for the entire simulation up to the current time.

ledger is an instance of Ledger which is maintaining the simulation’s state. This may be used to lookup the algorithm’s current portfolio values.

dt is a Timestamp which is the label of bar that has just completed.

session_ix is an int which is the index of the current trading session being run. This is provided to allow for efficient access to the daily returns through ledger.daily_returns_array[:session_ix + 1].

data_portal is an instance of DataPortal which is the metric’s interface to pricing data.

Defining New Metrics Sets

Users may use zipline.finance.metrics.register() to register a new metrics set. This may be used to decorate a function taking no arguments which returns a new set of metric object instances. For example:

from zipline.finance import metrics

@metrics.register('my-metrics-set')
def my_metrics_set():
    return {MyMetric(), MyOtherMetric(), ...}

This may be embedded in the user’s extension.py.

The reason that a metrics set is defined as a function which produces a set, instead of just a set, is that users may want to fetch external data or resources to construct their metrics. By putting this behind a callable, users do not need to fetch the resources when the metrics set is not being used.