Volatility-Scaled Trend Portfolio

Thesis (edge)

Take two managers with the same trend signal. One sizes every market with the same dollar amount, the other sizes so that every market contributes the same amount of risk. Over 20 years, their results will look nothing like each other, and the second one will almost always look better. That gap is not a signal edge. It is a construction edge, and it is available to anyone willing to do the plumbing carefully.

The reasoning is simple. A 10-year note future and a natural gas future are not comparable. Natural gas can move 5 percent in a day, the note might move 0.4 percent. If you hold the same notional in both, natural gas is your entire portfolio and the note is a rounding error. Dividing each position by that market's recent volatility puts them on a level field, so that your 40-market book actually behaves like 40 bets rather than 4 bets plus 36 decorations.

The second layer is volatility targeting at the portfolio level. Rather than accepting whatever volatility the market hands you, you scale total exposure up when markets are calm and down when they are wild, so the programme delivers a roughly stable risk level. Investors care about this because a fund that runs at 6 percent volatility one year and 25 percent the next is impossible to size in their own portfolio.

Where it works (regimes)

The sizing layer works everywhere. It is close to a free improvement, because it is correcting an obvious mismatch rather than predicting anything.

The volatility targeting layer works well when volatility is persistent and mean reverting, which it usually is. It works badly at a sudden shock, because the estimate is backward looking. Volatility explodes, you were still sized for the calm period, you take the hit, and only then do you de-lever. Everyone who runs this admits it. The point is that over many cycles the benefit of not being over-levered in the average calm-to-stormy transition outweighs the cost of being late to the sharp ones.

Signals

The trend signal itself is deliberately not the focus here. Assume you already have one that outputs, for each market, a direction and a strength between fully short and fully long. This entry is about what happens next.

• Volatility estimate: an exponentially weighted standard deviation of daily returns, with a half-life of roughly one to two months. Blend it with a multi-year average so that a single quiet month does not produce enormous leverage.
• Position size: signal strength divided by volatility, so equal risk per market.
• Sector budget: group markets into equities, bonds, short rates, FX, energy, metals and agriculture, and cap each sector's share of total risk.
• Portfolio scalar: one number that multiplies the whole book so that expected volatility, computed from the covariance matrix, lands on the target.

Portfolio construction

The correlation matrix is where the real work is. Trend books look diversified on paper and turn out to be one bet in practice, because trend signals across markets all end up leaning on the same macro theme. When the dollar strengthens, you are short six currencies, long the dollar index, short gold and short emerging markets. That is one position.

Use a shrunk correlation estimate rather than the raw sample matrix, which is noisy and unstable with 40 markets and a few hundred observations. Then use it to work out how much risk the book is really taking, and be prepared to discover that your effective number of independent bets is far smaller than your number of markets.

Finally, cap leverage in absolute terms. Volatility targeting will happily tell you to run 6 times leverage in a very quiet market. That is the moment the model is most confidently wrong.

Risk model

Two things must be modelled explicitly. First, correlation breakdown: run a scenario where all sector correlations jump toward one and see what your book loses. Second, volatility gaps: run a scenario where volatility triples overnight while you are still sized for yesterday.

A drawdown ladder is the simplest control that actually helps. Reduce the volatility target by a set fraction at each drawdown threshold, and restore it slowly. It costs performance in a V-shaped recovery, and it saves you in the drawdowns that keep going.

Costs & implementation

The construction layer generates its own turnover, separate from the signal. Every time the volatility estimate moves, position sizes move, even if the signal has not changed at all. If you rebalance every day to the exact target, a meaningful chunk of your return goes to brokers.

Use a no-trade band: only adjust a position when the target differs from the current position by more than some threshold. Slowing the volatility estimator also helps. The performance difference between a fast and a slow volatility estimate is usually small, but the cost difference is not.

Failure modes

• Believing volatility targeting protects against losses. It targets volatility, not losses, and trends can lose money at perfectly normal volatility.
• Using a raw, unshrunk covariance matrix and getting extreme, unstable position sizes.
• Letting leverage run unbounded in quiet markets.
• Rebalancing to target continuously and bleeding out in costs.
• Ignoring sector concentration and running a single macro bet under a diversified label.

Our Notes & Suggestions

Build this as layers and measure each one. Run the raw signal with equal notional sizing, then add inverse volatility sizing, then sector caps, then portfolio volatility targeting, then the drawdown ladder. Record the Sharpe ratio and the worst drawdown at each step. You will usually find that inverse volatility sizing gives the biggest single improvement and that most of the fancier layers add much less than you expect.

Be honest in the marketing. A programme that says it targets 10 percent volatility will still have months well outside that range. The target is an average, not a promise.