The most expensive sentence in systematic investing is “I am confident, so I will size it up.” Confidence is not a risk measurement. It is a feeling about a forecast, and it tends to be highest right before the environment that produced the forecast stops existing.

This note describes the sizing layer I use across multi-sleeve portfolios: a small set of regime states estimated from observable data, a volatility budget that changes with those states, and a de-grossing rule that reduces exposure on evidence of deterioration rather than on a forecast of it.

01The sizing problem

Equal-weighting across signals looks diversified and is not. If three sleeves are long the same liquidity beta, equal weights produce a portfolio whose true risk is concentrated in one factor. Inverse-volatility weighting improves the risk balance in calm markets and fails exactly when correlations converge, because the volatilities were estimated in a different world.

What is needed is a sizing rule that is (a) estimated only from data available at the time, (b) explicit about the regime it assumes, and (c) willing to hold less risk than the model says it could, because the model is known to be wrong occasionally and catastrophically.

02Classifying regime without hindsight

The classifier uses four observable dimensions, each computed on a rolling basis so that no future information enters:

  • Trend. Broad equity index versus its long moving average, plus the slope. Captures direction and persistence.
  • Volatility. Realized volatility percentile against its own three-year history. Captures the price of risk.
  • Correlation. Average pairwise correlation across sleeves, and its change. Captures whether diversification is currently working.
  • Stress. Credit spread direction and the drawdown of the portfolio itself. Captures the part of the world that equity prices sometimes ignore.

These are mapped into three states — expansion, contraction, shock — with hysteresis: moving into a riskier state requires more evidence than staying in it, and leaving a riskier state requires more evidence than entering. That asymmetry is deliberate. The cost of being late out of a shock is much larger than the cost of being early.

StateDefinition (summary)Risk budgetVol targetMax gross
ExpansionTrend positive, vol below median, correlation stableFull10%100%
ContractionTrend deteriorating or vol above medianReduced7%70%
ShockVol spike, correlation spike, or spread wideningDefensive4%35%

03The risk budget

The budget is expressed as a target portfolio volatility, not a target return. Each sleeve receives a share of that budget based on its own volatility and its marginal contribution to portfolio risk, which is recomputed whenever the correlation matrix moves materially. When a sleeve’s volatility doubles, its weight does not halve automatically — the marginal contribution is what matters, and that depends on what the rest of the book is doing.

Two hard constraints sit on top of the optimizer. A per-sleeve cap prevents a low-volatility sleeve from taking the entire budget. A cluster cap prevents correlated sleeves from acting as one position. Both are set from the strategy’s mandate, not from the data, because the data cannot tell you how much concentration you are willing to survive.

04The de-grossing ladder

Tactical de-risking usually fails for one of two reasons: it triggers so late that it sells the bottom, or it triggers so often that costs consume the benefit. The ladder addresses both by making the response graded and pre-committed.

  1. 01Level 0 — normal.Risk budget per the regime table. No discretionary override.
  2. 02Level 1 — trim.Portfolio drawdown exceeds 1.5 times the modeled expectation, or realized volatility exceeds target by 40%. Gross cut by a fixed fraction; no new entries in the weakest-scoring sleeve.
  3. 03Level 2 — halve.Two consecutive trigger observations, or a correlation spike into the top decile. Gross capped at half the normal budget; only the highest-conviction, lowest-correlation sleeves remain.
  4. 04Level 3 — flat.Shock state confirmed, or a drawdown beyond the mandate’s tolerance. Everything to cash except explicitly designated hedges. Re-entry requires the shock condition to clear and the trend to re-establish.

The ladder is deliberately mechanical. The decision to de-risk is made when the portfolio is calm, in writing, with the thresholds fixed. That is the only time it can be made well.

05Implementation sketch

The production version runs in Python on a daily schedule. The core is small: state estimation, budget allocation, and a constraint layer that refuses to trade outside pre-set limits. Below is the shape of it, simplified for readability.

def risk_budget(state, realized_vol, corr):
    base = {"expansion": 0.10, "contraction": 0.07, "shock": 0.04}[state]
    # scale the budget down when realized risk is running hot
    scale = min(1.0, base / max(realized_vol, 1e-4))
    # penalize the budget when diversification is failing
    if corr.avg_pairwise > corr.p90:
        scale *= 0.7
    return base * scale

def size_sleeves(scores, vols, cov, budget):
    raw = risk_parity_weights(vols, cov)          # marginal risk contribution
    raw = apply_caps(raw, sleeve_cap=0.35, cluster_cap=0.55)
    raw = raw * (budget / portfolio_vol(raw, cov)) # scale to budget
    return apply_trade_bands(raw, band=0.03)       # only trade if it matters

The important line is the last one. Trade bands are not an optimization detail; they are a statement that a 2% target change is not worth the spread. The strategy is judged on risk-adjusted return after costs, and most of the cost lives in small, unnecessary trades.

06The whipsaw budget

Every de-risking rule has a cost: sometimes it sells and the market immediately recovers. Rather than pretend otherwise, the ladder carries an explicit whipsaw budget — a maximum number of round trips per year that the rule is allowed to generate before it is considered too sensitive and must be retuned.

This changes the research question. Instead of asking “does de-risking improve the Sharpe ratio,” I ask “does it improve the left tail enough to justify the round trips it consumes.” That framing is what produces the hysteresis in the ladder: the model has to be slow to re-enter, not only quick to exit, because re-entry is where the round trips are actually spent.

07What this does not solve

Regime models are lagging by construction. They will be wrong at every turning point, and the ladder exists because they will be wrong. The realistic objective is not to avoid drawdowns but to ensure that the portfolio’s risk in the worst month is a decision that was made in advance, not a surprise.

Three limitations worth stating to anyone considering this framework: it cannot fix a bad signal, it cannot create diversification that is not there, and its parameters encode a view about how much pain the mandate can take. That last one is a governance question, not a quantitative one. It belongs to the allocator, not the model.

For discussion only. Not investment advice, not an offer to buy or sell any security, and not a representation of actual or expected performance.

Back to work