The safe portfolio guarantee that still permits a 20% shortfall

A prediction can point up, point down, or deserve no directional bet at all. MohammadpourAzari, Zarei, and Fazli make those three possibilities explicit: Follow the prediction, Reverseit, or transfer the decision to a fully invested, low-turnover Safe portfolio.

The paper is unusually disciplined about what happened next. A sign chosen on 24 development stocks failed on a frozen, disjoint 82-stock universe. Flipping the sign after seeing that result also failed to beat equal weight. The adaptive controller is therefore a story about containing the cost of untrusted advice—not discovering alpha.

The controller builds three implementable portfolios from the same market state. Safe ignores the prediction. Follow adds a centered long-short overlay in the prediction's direction. Reverse adds the opposite overlay. A probability vector says how much of the decision belongs to each action.

Matured rewards are log wealth relative to Safe, scaled into [−1, 1]. Because positions last five trading days, several cohorts overlap. A decision opened today cannot learn from an outcome that matures later today; the paper keeps those clocks separate.

Ordinary exponential weighting can drive a losing action's probability so close to zero that a later polarity change is slow to repair. Fixed-Share first rewards the actions exponentially, then redistributes a fraction of probability uniformly. That leak is a recovery channel, not permission to trade.

For each directional action, the implementation keeps the latest 50 matured rewards. After at least 25 observations, it accepts the action only if the sample mean minus 1.645 standard errors is positive. The paper calls this a nominal one-sided normal bound because overlapping cohorts make consecutive rewards dependent.

def gate_probabilities(raw, accept_follow, accept_reverse):
    follow = raw[1] if accept_follow else 0.0
    reverse = raw[2] if accept_reverse else 0.0
    safe = 1.0 - follow - reverse
    return [safe, follow, reverse]

The tested TypeScript function runs the chart calculations; Python and C++ are line-for-line translations. Rejected Follow or Reverse mass is moved to Safe. The learner underneath is unchanged, so later matured evidence can revive the action.

CAGR is compound annual growth rate; Sharpe divides excess return by volatility; maximum drawdown is the deepest peak-to-trough loss; turnover measures how much portfolio weight is traded. On the frozen universe, the pre-specified Reverse strategies did not beat daily equal weight. The conservative anchor stayed close with much less turnover, but closeness is not alpha.

On the 24-stock development universe, Safe matched equal weight's broad return profile while cutting average daily turnover from 1.383% to 0.104%—a core-computed reduction of 92.5%. Certified execution was identical to Safe because the directional lower bound never turned positive. The prediction layer was available, but never earned permission to trade.

After the frozen Reverse result failed, Follow was better under both stored targets. It still remained below equal weight. Selecting the better sign from the same holdout would turn diagnosis into post-hoc model selection, so the paper labels these rows exploratory and leaves the confirmatory verdict untouched.

First, replace the nominal normal gate with a dependence-aware bound or a block-bootstrap gate and report how much active mass changes. The current rule uses overlapping five-day rewards, so “1.645 standard errors” is a configuration choice rather than a finite-sample coverage certificate.

Next, move the frozen test forward in time as well as across stocks, stress nonlinear transaction costs and liquidity capacity, and compare the signed controller with a simple probability model whose calibration can be inspected on the logistic regressionpage. A safer wrapper cannot manufacture a forecast edge; it can only ration how much damage a missing edge is allowed to cause.