Momentum strategies have a dirty secret. They work until the exact moment they stop working, and when they stop, the losses come fast. A standard cross-sectional momentum portfolio lost 42.58% in 2008. The winners it had loaded up on became the fastest fallers. I have run enough trend-following screens to know the pattern: the stocks with the prettiest equity curves going into a crash are often the ones that hurt you worst coming out.
An April 2026 paper by Arya Chakraborty and Randhir Singh (arXiv:2604.09060) directly attacks this problem. Their framework, called AEGIS (Adaptive Equity Generation and Immunisation System), adds two layers on top of standard volatility-gated momentum selection. The first gates assets by volatility-adjusted returns. The second enforces structural diversification through a minimax correlation filter. Over a 20-year walk-forward backtest from 2006 to 2025, the result was a 15.41% compound annual growth rate with a maximum drawdown of 28.89%. For context, the S&P 500 drew down over 50% in that same window, and a standard momentum strategy lost 42.58% in 2008 alone.
The concept is directly relevant if you use any form of network momentum or lead-lag signals in your trend-following, because the same correlation clustering that destroys momentum portfolios also poisons lead-lag networks during crises.
Why Standard Momentum Crashes
The academic term is the “Winner’s Curse.” Standard momentum ranks stocks by trailing returns over some lookback period, typically 12 months. It buys the top performers and holds. The problem: those top performers tend to cluster in the same sectors, carry similar volatility profiles, and move together when sentiment shifts.
When a bear market hits, correlations spike toward 1.0. The portfolio that looked diversified across 30 names turns out to be one big bet on the same risk factor. Daniel and Moskowitz documented this in their 2016 paper on momentum crashes. The losses are not gradual. They are concentrated in a few weeks as crowded trades unwind simultaneously.
I have watched this play out in real time with swing screens. You build a watchlist of 20 stocks with strong regime-factor profiles, and three weeks later half of them gap down on the same macro headline. The problem is not the momentum signal. The problem is the correlation structure underneath it.
Volatility-Gated Momentum as the First Filter
The AEGIS framework starts with what it calls a Volatility-Adjusted Momentum (VAM) signal. Instead of ranking stocks purely by cumulative return, VAM divides each stock’s trailing return by its trailing volatility. The formula normalises momentum by the risk taken to generate it:
VAM_i = \frac{R_i}{\sigma_i}
A stock that returned 40% with 15% annualised volatility scores higher than one that returned 50% with 30% volatility. The first moved efficiently. The second was along for a wild ride that could reverse just as violently.
The paper uses a 12-month lookback for the selection signal and ranks the entire S&P 500 universe. Only the top three VAM-ranked stocks become the “anchor triad” for the portfolio. This is the aggressive core. Everything else gets built around it to dampen risk.
Most traders who run momentum screens skip this normalisation step. They sort by absolute return and pick the top names. That approach systematically overweights high-beta stocks near cycle peaks. If you are building any kind of multi-factor composite screen, adding a volatility denominator to your momentum score is the single cheapest improvement you can make. It does not require new data. It just reweights what you already have.
Minimax Correlation Diversification
The second layer is where the framework gets interesting. After selecting the anchor triad, AEGIS needs to fill 47 more positions (for a total basket of 50). It does not pick them by return or sector. It picks them by how little they correlate with what is already in the basket.
The algorithm is greedy and recursive. For each candidate stock, it calculates the maximum absolute pairwise correlation against every stock already in the portfolio:
\rho_{max}(c) = \max_{b \in B} |corr(r_c, r_b)|
Then it selects the candidate with the lowest maximum correlation. The new stock is added to the basket. The correlation matrix updates. The process repeats until the basket reaches 50 names.
There is one critical gate before any candidate enters the pool: its cumulative return over the lookback period must be positive. This prevents the algorithm from selecting distressed or declining stocks simply because they happen to be uncorrelated with everything else. Without this gate, the algorithm would fill the basket with value traps and broken names that are uncorrelated precisely because they are dying.
This is the piece most correlation-based diversification methods get wrong. They optimise for low correlation without checking whether the assets they are adding actually have positive drift. AEGIS forces both conditions: positive momentum and low correlation.
The Allocation Engine and Sortino Optimisation
With 50 stocks selected, the framework allocates capital using a constrained optimisation engine. It maximises the Sortino ratio, not the Sharpe ratio. The distinction matters. The Sharpe ratio penalises all volatility equally, including upside moves. The Sortino ratio only penalises downside deviation below a target return:
Sortino = \frac{R_p - R_f}{DD_{ann}}
where DD_{ann} is the annualised lower partial moment of degree 2, using a 4% risk-free hurdle rate.
The optimiser runs on a rolling 3-month lookback window. Every month, it recalculates weights based on the most recent three months of returns and covariance. Weights are frozen and applied to the next month, then recalculated. No position can exceed 5% of the portfolio. No short selling. The portfolio must be fully invested.
The 3-month window was a deliberate choice, not the result of curve-fitting. A 12-month allocation lookback produced higher absolute returns (18.39% CAGR) but much higher volatility (23.61% annualised). The 3-month window compressed volatility to 16.44% because it forces the optimiser to react to recent regime changes. In their robustness sweep, the 3-month/50-stock configuration dominated every other combination on a risk-adjusted basis.
If you have read about walk-forward analysis on this site, this is the same principle. Train on the recent past. Apply to the next unseen period. Never optimise on the full dataset.
How AEGIS Performed During Crises
The 20-year backtest covers three major market dislocations: the 2008 Global Financial Crisis, the March 2020 COVID crash, and the 2022 rate-tightening selloff. The results across all three are what make this framework worth studying.
In 2008, the S&P 500 suffered a maximum drawdown exceeding 50%. The NASDAQ fell over 41%. A standard cross-sectional momentum strategy lost 42.58% for the year. AEGIS posted a net loss of 20.94% with a maximum drawdown of 28.00%. Still painful, but roughly half the damage of unprotected momentum. The minimax correlation layer forced the portfolio away from the concentrated financial and cyclical bets that destroyed standard momentum screens.
In 2022, the framework performed even better on a relative basis. While the NASDAQ dropped 33% and a stock-bond risk parity strategy collapsed 26.72% as bonds and equities fell together, AEGIS limited its calendar-year loss to 5.10%. The rolling 3-month allocation window detected the correlation shift between equities and bonds in real time and rotated capital into structurally orthogonal positions.
That 2022 result is the one I find most interesting for active traders. Risk parity, the strategy that “always works,” failed precisely because the historical inverse correlation between stocks and bonds broke down during an inflationary tightening. AEGIS does not rely on static asset-class correlations. It recalculates rolling covariance every month. When a relationship breaks, it adapts within one or two rebalancing cycles.
What the Numbers Actually Show
Over the full 20-year period, the framework compounded $10,000 into $176,302 net of transaction costs. The S&P 500 reached $54,838 over the same period. The NASDAQ-100 reached approximately $176,000, nearly identical terminal wealth, but through a much more volatile path with drawdowns exceeding 40% twice.
The key statistics from the paper: 15.41% annualised net return. 16.44% annualised volatility. Maximum drawdown of 28.89%. Monthly win rate of 68.8%. Profitable in 18 of 20 calendar years. Outlier-adjusted Sortino ratio of 1.72 (after removing the two best years where near-zero drawdowns inflated the ratio to extreme values). Transaction friction averaged less than 1% per year, totalling 15.58% over the full period.
For the robustness-minded: the paper tested basket sizes of 25, 50, and 75, and allocation lookbacks of 3, 6, and 12 months. The 25-stock basket suffered drawdowns exceeding 50% because it was under-diversified. The 75-stock basket diluted the momentum premium and compressed returns to the 11-13% CAGR range. The 50-stock basket with 3-month allocation was the sweet spot.
Where the Framework Falls Short
The paper is honest about its scope, but there are practical limitations worth flagging. The backtest uses monthly rebalancing on the full S&P 500 universe. For a retail trader or small fund, executing a 50-stock portfolio rebalanced monthly is not trivial. Slippage on the smaller names in the basket would likely exceed the 0.1% round-trip friction assumed in the paper.
The 2009 result is telling. After the 2008 crash, AEGIS returned only 2.76% while the S&P 500 surged over 26%. The VAM filter, by penalising volatility, kept the portfolio out of the high-beta recovery stocks that led the rebound. This is the cost of crash protection: you miss the sharpest bounces because the same volatile names you avoided on the way down are the ones that snap back hardest.
Survivorship bias is partially addressed. The authors injected delisted stocks back into the universe to prevent the backtest from benefiting from hindsight on which companies survived. The Monte Carlo stress testing approach we have covered separately would be a useful complement: simulate how the framework behaves when delisted stocks are more frequent or when correlation spikes are more severe than historical experience.
The paper also does not address sector concentration within the anchor triad. If the three highest-VAM stocks happen to be in the same sector, the diversification layer only kicks in for positions 4 through 50. The core bet remains concentrated. A sector-diversity constraint on the anchor triad would be a natural extension.
Practical Takeaways for Trend Followers
You do not need to build the full AEGIS pipeline to use its core ideas. Three principles from the framework translate directly to any momentum or volatility-regime position sizing approach.
First, normalise your momentum signal by volatility. Divide trailing return by trailing standard deviation before ranking. This single adjustment filters out the “rising on borrowed risk” stocks that look strong but are structurally fragile. I started doing this with my own screens after reading the paper, and the immediate effect was removing several high-beta tech names that had dominated pure-return rankings.
Second, check the correlation structure of your holdings, not just their individual signals. A portfolio of five strong momentum stocks that all correlate at 0.85 is functionally a one-stock portfolio. The minimax approach, picking each new addition to minimise the worst-case correlation against existing holdings, is a practical method that does not require a full optimisation engine. A spreadsheet and a correlation matrix will get you most of the way there.
Third, use a short allocation lookback. The 3-month window beat the 12-month window on risk-adjusted returns because it adapts faster to regime changes. If you are rebalancing a swing portfolio monthly or bi-weekly, your allocation weights should be based on recent data, not the last year. Stale covariance estimates are what killed risk parity in 2022.
Where Volatility Gating Fits in Your Process
Volatility-gated momentum is not a replacement for your entry signals or your stop-loss rules. It sits one level above: at the portfolio construction layer. It answers the question “which stocks should I be trading” before you ask “when should I enter this stock.”
If you currently run a drawdown-quality momentum filter, adding a volatility gate is the logical next step. The drawdown filter tells you which stocks recover cleanly from pullbacks. The volatility gate tells you which stocks earned their momentum efficiently. Together, they narrow the universe to names that trend well and trend smoothly.
The AEGIS paper demonstrates that even a relatively simple set of rules, volatility-adjusted ranking, recursive correlation filtering, and Sortino-optimised allocation, can cut the worst-case drawdown of a momentum strategy roughly in half while preserving most of the upside. That is the kind of asymmetry worth engineering into any trend-following system.
Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.