Multi-Asset Momentum and Trend-Following in 2016

Most multi-asset portfolios use static allocation: a fixed percentage to equities, a fixed percentage to bonds, rebalanced on a schedule. The TrendFolios paper asks whether replacing that static weight with dynamic signal-driven inclusion decisions across all three buckets simultaneously produces better risk-adjusted returns than a traditional 60/40 over a long period. The test runs from December 1997 through December 2023, which is long enough to include the dot-com collapse, the 2008 financial crisis, the European debt crisis, and 2020. The result is mostly yes, but with a pattern in the data that is more interesting than the headline number.

Two Signals, One Decision

The framework uses two independent signals for each of its 21 ETF positions. The first is a momentum signal: a line-based measure of relative returns across timeframes spanning daily to annual, adjusted for relative volatility and spread between instruments. The second is a trend-following signal: a curve-based measure of price trajectory. These are not the same thing. A strong recent return is momentum. A price consistently above a moving average is trend. An asset can have one without the other.

The inclusion decision uses a majority vote. If the two signals agree, the position is included. If they disagree, it is not. This is not a blended score where partial agreement produces a partial position. The rule is binary: agree, you are in; disagree, you are out. The practical effect is that the framework sits out of positions during periods when momentum and trend signals are pulling in opposite directions. For anyone who has traded through a whipsaw market where short-term momentum flips while longer-term trends are still intact, or vice versa, the logic is straightforward: conflicting signals are a reason to reduce exposure, not average it.

Why the Weighting Method Matters

Position sizing uses inverse volatility weighting, but the paper makes a specific choice about how to measure volatility: tracking error relative to a benchmark, not standard deviation of the instrument itself. The distinction is meaningful. Standard deviation measures how much an asset moves in absolute terms. Tracking error measures how much it deviates from its benchmark. In a multi-asset portfolio where each sub-universe has its own benchmark, tracking error captures benchmark-relative risk rather than raw price volatility. A highly volatile commodity ETF that moves in lockstep with its commodity index has low tracking error and gets a larger allocation than its absolute volatility would suggest. A bond ETF that drifts significantly from its benchmark gets penalized more than a raw vol calculation would indicate.

The Universe and Data

The 21 ETFs are divided into three equal sub-strategies of seven ETFs each: equity factors (large and small cap growth and value, developed and emerging markets), fixed income (inflation-protected, duration-varied, and credit instruments across developed and emerging markets), and alternatives (private equity, credit, real estate, energy, agricultural, metals, and precious metals commodities). The model uses only ETFs with live price history at each historical point, expanding the universe as new instruments become available. No synthesized pre-ETF data is used. Rebalancing is bi-weekly. The 0.55% annual management fee is deducted from all reported net figures. The RealTest ETF Rotate Monthly Rebalance Strategy operates on the same class of problem in RealTest: signal-driven ETF selection with configurable rebalancing frequency and cost modeling built into the strategy code.

Results: Where It Works and Where It Struggled

The headline result for the Moderate Portfolio (60% equity, 30% fixed income, 10% alternatives) is 7.94% annually gross over 26 years, 3.14% above a traditional 60/40 portfolio, with an information ratio of 0.33. Sharpe ratios ranged from 0.29 to 0.72 across sub-strategies. Worst year was 2008 at -38.73%, best year was 2009 at +36.58%.

The more informative result is in the sub-strategy breakdown. Fixed income generated 6.48% annual excess return above the Bloomberg Aggregate with an information ratio of 0.34. Alternatives generated 5.29% excess return. Equities generated 2.26% excess return above the S&P 500 with an information ratio of 0.27, which is the weakest of the three, and the authors note it underperformed in 1-3 year rolling windows following 2008.

That equity underperformance pattern is the most practically relevant finding in the paper. Momentum and trend signals depend on return dispersion: assets need to separate into persistent winners and losers for the signals to generate consistent alpha. When central banks suppressed volatility across equity markets through near-zero rates and quantitative easing after 2008, that dispersion compressed. Fixed income and alternatives were less affected by this because central bank policy created, rather than eliminated, differentiation between duration, credit, and commodity-linked assets. The paper cites Lo (2016) on signal crowding as contributing context, though the mechanism is not formally tested.

What the In-Sample Caveat Means in Practice

The study is in-sample throughout. The same historical data used to develop the framework is used to evaluate it. This is the paper's most significant methodological limitation, and the authors acknowledge it directly, noting out-of-sample validation is pending.

For a systematic trader evaluating whether to build something similar, this means the reported numbers are an upper bound on what out-of-sample performance might look like, not a baseline. The information ratios of 0.27-0.34 across sub-strategies are plausible for a systematic multi-asset approach, but they are measured on the same data the model was tuned on. Parameter choices for lookback windows, signal thresholds, and the tracking-error weighting calculation are not disclosed, which makes direct replication impossible and prevents independent assessment of how sensitive the results are to those choices.

In RealTest, the appropriate response is to implement a version of the framework, hold out the most recent 5 years as a test set, and evaluate whether the information ratio and excess return pattern hold. 

Building a Version in RealTest

The paper withholds specific parameters, but the design is specific enough to implement a comparable framework:

• Build three ETF sub-universes in RealTest using Norgate Data. Use tickers that have been trading since at least 2003 to get a reasonable history before the 2008 event
• For the momentum signal, relative strength ranking over multiple lookback periods is a reasonable starting point. Try 1-month, 3-month, and 6-month equally weighted, excluding the most recent 20 trading days to avoid reversal contamination
• For the trend signal, price relative to a 200-day exponential moving average is the simplest version. More complex options include linear regression slope or a dual moving average crossover
• Implement the majority vote as: include the position only if both signals independently classify the ETF as a buy. If either says no, the position is zero
• For volatility-based sizing, rolling 60-day ATR relative to the ETF's benchmark ETF as a tracking error proxy is a practical approximation. Inverse-weight by this measure across included positions within each sub-universe
• Set the rebalancing calendar to every 10 trading days (bi-weekly)
• Test from 2003 at earliest. Do not start from 2009 because it begins with the recovery from the worst drawdown and will produce inflated results
• Hold out 2019-2023 as your test set and only look at those results after finalizing the parameters on 2003-2018

Given that the paper does not report out-of-sample validation, any replication that produces in-sample numbers matching the paper's figures should be treated as a starting point rather than a validated result until the held-out period is evaluated.

Limitations

No out-of-sample validation. This is stated by the authors and is the primary reason to treat the headline figures cautiously rather than as a performance forecast.

The equity sub-strategy underperformed post-2008. The paper attributes this to monetary policy, which is a plausible explanation, but if that explanation is correct, the question of whether that compression is permanent or temporary is open. The fixed income and alternatives results may face analogous headwinds in a rising-rate environment that is structurally different from 2009-2021.

The 0.55% annual fee represents institutional asset management pricing. At bi-weekly rebalancing across 21 positions, the transaction cost contribution to that number is non-trivial. A self-managed RealTest implementation would have lower management fees but potentially higher per-trade commissions depending on position sizes, making direct comparison to the net figures approximate.

The ETF universe is limited to US-listed instruments. The paper does not test whether the same signal logic applied to non-US equities, European fixed income, or Asian commodity markets would produce comparable excess returns.

Citation: Lu, J., Rojas, R.R., Yeung, F.C., and Convery, P.D. (2025). TrendFolios: A Portfolio Construction Framework for Utilizing Momentum and Trend-Following In a Multi-Asset Portfolio. arXiv preprint arXiv:2506.09330. https://arxiv.org/abs/2506.09330