Failure and Fallbacks

This tutorial introduces the optimization parameters fallback and raise_on_failure.

Optimization can sometimes fail during a given rebalancing. For example, a convex mean-variance problem with strict risk or sector constraints may become infeasible on specific dates. Such failures must be handled explicitly depending on the use case (production vs. research).

Fallback

The fallback parameter lets you define an estimator, or a list of estimators, to try in order when the primary optimization raises an error during fit. Alternatively, you can use "previous_weights" to reuse the last valid allocation.

Each attempt is recorded in fallback_chain_, and the successful estimator is available through fallback_.

This mechanism is essential in automated pipelines, ensuring that optimization failures never halt production runs while preserving full reproducibility and traceability. Beyond safeguarding workflows, it can also be used to deliberately relax constraints in a controlled manner when strict convergence cannot be achieved.

Raise on Failure

In research, cross-validation and hyperparameter tuning (e.g. walk-forward, multiple randomized cross-validation), it’s often useful to let all runs complete while keeping a full record of failures instead of stopping on the first failed rebalancing.

• Set raise_on_failure=True (default) to fail fast. This is useful in production when the primary optimization or the fallback cascade is expected to succeed.

• Set raise_on_failure=False to continue uninterrupted. This is useful in research and cross-validation. When a failure occurs, predict returns a FailedPortfolio (think of it as an augmented NaN) that carries diagnostics such as optimization_error and fallback_chain, while remaining API-compatible with downstream analytics.

Data and Setup

Load the S&P 500 dataset and split into train/test.

import pandas as pd
from plotly.io import show
from sklearn.model_selection import train_test_split
from sklearn.utils.validation import validate_data

from skfolio.datasets import load_sp500_dataset
from skfolio.model_selection import WalkForward, cross_val_predict
from skfolio.optimization import BaseOptimization, EqualWeighted, MeanRisk
from skfolio.preprocessing import prices_to_returns
from skfolio.typing import Fallback, MultiInput
from skfolio.utils.stats import rand_weights

# Load S&P 500 dataset and split train/test
prices = load_sp500_dataset()
prices = prices["2010":]
X = prices_to_returns(prices)
X_train, X_test = train_test_split(X, test_size=0.33, shuffle=False)

Fallback

Let’s start with a simple example. The primary model is a minimum-variance optimization made intentionally infeasible (the assets’ minimum weights are set to 10%, which exceeds the feasible upper bound of 1/n_assets = 5%). As a fallback, we provide a feasible minimum-variance model with a 2% minimum weight constraint:

model = MeanRisk(
    min_weights=0.1,  # intentionally infeasible
    fallback=MeanRisk(min_weights=0.02),  # feasible fallback
)
model.fit(X_train)
print(model.weights_)

[0.02000143 0.02000011 0.02000017 0.02000038 0.02000046 0.02000042
 0.02000085 0.12397314 0.02000024 0.09473377 0.02004185 0.02000133
 0.02000057 0.17901455 0.0200018  0.17643479 0.02000026 0.02000095
 0.12579215 0.02000078]

Diagnostics

Let’s retrieve the fitted fallback that produced the final result:

print(model.fallback_)

MeanRisk(min_weights=0.02)

Let’s display the sequence of attempts and their outcomes:

print(model.fallback_chain_)

[('MeanRisk(fallback=MeanRisk(min_weights=0.02), min_weights=0.1)', "Solver 'CLARABEL' failed. Try another solver, or solve with solver_params=dict(verbose=True) for more information"), ('MeanRisk(min_weights=0.02)', 'success')]

The fallback audit trail is also propagated to the predicted portfolio:

portfolio = model.predict(X_test)
assert portfolio.fallback_chain == model.fallback_chain_

Multiple fallbacks

We can also provide a list of fallbacks to be tried in order, including “previous_weights” as a terminal safety net:

model = MeanRisk(
    min_weights=0.1,
    previous_weights={
        "AAPL": 0.4,
        "AMD": 0.2,
        "UNH": 0.4,
    },  # any missing assets default to 0
    fallback=[
        MeanRisk(min_weights=0.02),
        MeanRisk(min_weights=0.01),
        EqualWeighted(),
        "previous_weights",
    ],
)

Chaining

We can also nest fallbacks. The chain is evaluated depth-first from the primary estimator to the first successful fallback, recording each attempt in fallback_chain_. This is equivalent to providing an ordered list:

model = MeanRisk(
    min_weights=0.1,
    fallback=MeanRisk(
        min_weights=0.02,
        fallback=MeanRisk(
            min_weights=0.01,
            fallback=EqualWeighted(),
        ),
    ),
)

Fallback in cross-validation

Fallback behavior is fully preserved in cross-validation.

When using cross_val_predict, all diagnostics (e.g., fallback chains and errors) are propagated to the resulting portfolios in the MultiPeriodPortfolio:

• Each individual Portfolio (or FailedPortfolio) produced during rebalancing carries its own fallback_chain and optimization_error.

• Global counts and statistics (e.g., the number of portfolios that required a fallback) are available through summary attributes such as n_fallback_portfolios and n_failed_portfolios.

• The summary() method consolidates performance and diagnostic information across all rebalances.

model = MeanRisk(min_weights=0.1, fallback=MeanRisk(min_weights=0.02))

# Rebalance semiannually on the third Friday (WOM-3FRI), training on the prior 12 months
walk_forward = WalkForward(test_size=6, train_size=12, freq="WOM-3FRI")

pred = cross_val_predict(model, X, cv=walk_forward)

Let’s retrieve the fallback chain of the first portfolio:

print(pred[0].fallback_chain)

[('MeanRisk(fallback=MeanRisk(min_weights=0.02), min_weights=0.1)', "Solver 'CLARABEL' failed. Try another solver, or solve with solver_params=dict(verbose=True) for more information"), ('MeanRisk(min_weights=0.02)', 'success')]

Let’s print the number of portfolios in MultiPeriodPortfolio where a fallback was used:

print(pred.n_fallback_portfolios)

23

Finally, let’s display the last four rows of the MultiPeriodPortfolio summary, which contain the fallback statistics:

print(pred.summary().iloc[-4:])

Avg nb of Assets per Portfolio    20.0
Number of Portfolios                23
Number of Failed Portfolios          0
Number of Fallback Portfolios       23
dtype: object

Failure handling

In this section, we show how to handle optimization failures using the raise_on_failure parameter. As an example, we create a custom optimization that intentionally fails during fit when the first date of the input window falls on an even day of the month, or when always_fail=True.

class CustomOptimization(BaseOptimization):
    """Dummy optimization that intentionally fails during `fit` when the first
    date of the input window is an even day-of-month, or when `always_fail=True`."""

    def __init__(
        self,
        always_fail: bool = False,
        portfolio_params: dict | None = None,
        fallback: Fallback = None,
        previous_weights: MultiInput | None = None,
        raise_on_failure: bool = True,
    ):
        super().__init__(
            portfolio_params=portfolio_params,
            fallback=fallback,
            raise_on_failure=raise_on_failure,
            previous_weights=previous_weights,
        )
        self.always_fail = always_fail

    def fit(self, X: pd.DataFrame, y=None):
        validate_data(self, X)
        # Fail when first observation date has an even day-of-month, or always.
        if self.always_fail:
            raise RuntimeError("Forced failure")
        first_day = X.index[0].day
        if first_day % 2 == 0:
            raise RuntimeError("Forced failure (even-start window)")
        n_assets = X.shape[1]
        self.weights_ = rand_weights(n_assets)
        return self

By default, as with all scikit-learn estimators, failures raise an error during fit:

model = CustomOptimization(always_fail=True)
try:
    model.fit(X_train)
except RuntimeError as err:
    print(err)

Forced failure

By setting raise_on_failure=False, a warning is emitted instead of raising an error, and weights_ are set to None, with the error message stored in error_:

model = CustomOptimization(always_fail=True, raise_on_failure=False)
model.fit(X_train)
print(model.weights_)
print(model.error_)

None
Forced failure

In this case, calling predict will return a FailedPortfolio carrying the audit trail in optimization_error and fallback_chain (if any fallbacks occurred).

portfolio = model.predict(X_test)
print(portfolio)
print(portfolio.optimization_error)

<FailedPortfolio CustomOptimization>
Forced failure

Setting raise_on_failure=False is useful for cross-validation and hyperparameter tuning as it allows all runs to complete without stopping at the first rebalancing failure. Let’s instantiate our custom optimization and run a walk-forward analysis where failures occur deterministically on even-start windows:

model = CustomOptimization(raise_on_failure=False)
pred = cross_val_predict(model, X, cv=walk_forward)

cross_val_predict completed without interruption. The resulting MultiPeriodPortfolio is composed of both Portfolio and FailedPortfolio objects:

print(pred.portfolios)

[<Portfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <FailedPortfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>, <Portfolio CustomOptimization>]

Let’s print the number of failed portfolios:

print(pred.n_failed_portfolios)

9

Even though MultiPeriodPortfolio contains failed portfolios, all statistics and plots still work properly. This is because FailedPortfolio is designed to behave like non-propagating NaNs:

print(pred.summary())

Mean                                     0.051%
Annualized Mean                          12.93%
Variance                                0.0069%
Annualized Variance                       1.73%
Semi-Variance                           0.0036%
Annualized Semi-Variance                  0.92%
Standard Deviation                        0.83%
Annualized Standard Deviation            13.17%
Semi-Deviation                            0.60%
Annualized Semi-Deviation                 9.57%
Mean Absolute Deviation                   0.61%
CVaR at 95%                               1.97%
EVaR at 95%                               2.44%
Worst Realization                         3.80%
CDaR at 95%                              12.14%
MAX Drawdown                             25.17%
Average Drawdown                          2.72%
EDaR at 95%                              16.33%
First Lower Partial Moment                0.30%
Ulcer Index                               0.042
Gini Mean Difference                      0.89%
Value at Risk at 95%                      1.39%
Drawdown at Risk at 95%                   8.32%
Entropic Risk Measure at 95%               3.00
Fourth Central Moment                 0.000003%
Fourth Lower Partial Moment           0.000001%
Skew                                    -13.38%
Kurtosis                                589.10%
Sharpe Ratio                              0.062
Annualized Sharpe Ratio                    0.98
Sortino Ratio                             0.085
Annualized Sortino Ratio                   1.35
Mean Absolute Deviation Ratio             0.084
First Lower Partial Moment Ratio           0.17
Value at Risk Ratio at 95%                0.037
CVaR Ratio at 95%                         0.026
Entropic Risk Measure Ratio at 95%      0.00017
EVaR Ratio at 95%                         0.021
Worst Realization Ratio                   0.013
Drawdown at Risk Ratio at 95%            0.0062
CDaR Ratio at 95%                        0.0042
Calmar Ratio                             0.0020
Average Drawdown Ratio                    0.019
EDaR Ratio at 95%                        0.0031
Ulcer Index Ratio                         0.012
Gini Mean Difference Ratio                0.058
Avg nb of Assets per Portfolio             20.0
Number of Portfolios                         23
Number of Failed Portfolios                   9
Number of Fallback Portfolios                 0
dtype: object

As shown below, MultiPeriodPortfolio plots gracefully handle FailedPortfolio instances; for cumulative returns, these appear as gaps corresponding to failed periods:

fig = pred.plot_cumulative_returns()
show(fig)

Finally, let’s inspect the first failed portfolio:

failed_ptf = pred.failed_portfolios[0]
print(failed_ptf.optimization_error)

Forced failure (even-start window)

To replay the optimization on the failed period, we can run:

# model.fit(failed_ptf.X)