add early stopping for hyperopt

docs/hyperopt.md

@@ -490,6 +490,8 @@ freqtrade hyperopt --config config.json --hyperopt-loss <hyperoptlossname> --str
 ```
 
 The `-e` option will set how many evaluations hyperopt will do. Since hyperopt uses Bayesian search, running too many epochs at once may not produce greater results. Experience has shown that best results are usually not improving much after 500-1000 epochs.  
+The `-es` option will set ofter how many batches of evaluations with no improvements hyperopt will stop. A good value is 20-30% of the total epochs. Early stop is by default disabled (`-es=0`)
+
 Doing multiple runs (executions) with a few 1000 epochs and different random state will most likely produce different results.
 
 The `--spaces all` option determines that all possible parameters should be optimized. Possibilities are listed below.

freqtrade/commands/arguments.py

@@ -78,6 +78,7 @@ ARGS_HYPEROPT = [
     "disableparamexport",
     "hyperopt_ignore_missing_space",
     "analyze_per_epoch",
+    "early_stop",
 ]
 
 ARGS_EDGE = [*ARGS_COMMON_OPTIMIZE, "stoploss_range"]

freqtrade/commands/cli_options.py

@@ -262,6 +262,14 @@ AVAILABLE_CLI_OPTIONS = {
         metavar="INT",
         default=constants.HYPEROPT_EPOCH,
     ),
+    "early_stop": Arg(
+        "-es",
+        "--early-stop",
+        help="Early stop hyperopt if no improvement after (default: %(default)d) epochs.",
+        type=check_int_positive,
+        metavar="INT",
+        default=0,  # 0 to disable by default
+    ),
     "spaces": Arg(
         "--spaces",
         help="Specify which parameters to hyperopt. Space-separated list.",

freqtrade/configuration/configuration.py

@@ -334,6 +334,11 @@ class Configuration:
             ("print_all", "Parameter --print-all detected ..."),
         ]
         self._args_to_config_loop(config, configurations)
+        if self.args.get("early_stop", 0) > 0:
+            logger.info(
+                f"Parameter --early-stop detected ... Will early stop hyperopt if no improvement "
+                f"after {self.args.get('early_stop')} epochs ..."
+            )
 
         configurations = [
             ("print_json", "Parameter --print-json detected ..."),

freqtrade/optimize/hyperopt/hyperopt.py

@@ -317,6 +317,13 @@ class Hyperopt:
                         logging_mp_handle(log_queue)
                         gc.collect()
 
+                        if (
+                            self.hyperopter.es_batches > 0
+                            and self.hyperopter.es_terminator.should_terminate(self.opt)
+                        ):
+                            logger.info(f"Early stopping after {(i + 1) * jobs} epochs")
+                            break
+
         except KeyboardInterrupt:
             print("User interrupted..")
 

freqtrade/optimize/hyperopt/hyperopt_optimizer.py

@@ -14,6 +14,7 @@ import optuna
 from joblib import delayed, dump, load, wrap_non_picklable_objects
 from joblib.externals import cloudpickle
 from optuna.exceptions import ExperimentalWarning
+from optuna.terminator import BestValueStagnationEvaluator, Terminator
 from pandas import DataFrame
 
 from freqtrade.constants import DATETIME_PRINT_FORMAT, Config
@@ -104,6 +105,11 @@ class HyperOptimizer:
 
         self.market_change = 0.0
 
+        self.es_epochs = config.get("early_stop", 0)
+        self.es_batches = self.es_epochs // config.get("hyperopt_jobs", 1)
+        if self.es_epochs > 0 and self.es_epochs < 0.2 * config.get("epochs", 0):
+            logger.warning(f"Easly stop epochs {self.es_epochs} lower than 20% of total epochs")
+
         if HyperoptTools.has_space(self.config, "sell"):
             # Make sure use_exit_signal is enabled
             self.config["use_exit_signal"] = True
@@ -424,6 +430,11 @@ class HyperOptimizer:
         else:
             sampler = o_sampler
 
+        if self.es_batches > 0:
+            with warnings.catch_warnings():
+                warnings.filterwarnings(action="ignore", category=ExperimentalWarning)
+                self.es_terminator = Terminator(BestValueStagnationEvaluator(self.es_batches))
+
         logger.info(f"Using optuna sampler {o_sampler}.")
         return optuna.create_study(sampler=sampler, direction="minimize")