Roy’s Safety-First Criterion
Table of Contents
A portfolio can earn a positive return and still leave an investor short of a goal. Roy’s safety-first criterion makes that minimum acceptable return the focus of portfolio selection. With a common investment horizon, it compares the chance that each portfolio will miss the chosen floor. The safety-first ratio provides a convenient ranking under a normal return model, but its name does not promise principal protection.
What Is Roy’s Safety-First Criterion?
The criterion selects the portfolio with the lowest probability of earning less than a specified minimum return. That event is called shortfall. Its definition depends on both the return floor and the investment horizon.[1]
A 0% floor asks about losing nominal principal over the period. A 6% floor asks about failing to earn 6%, including outcomes that still produce a gain. Those are different objectives and can lead to different portfolio choices.
For example, suppose an investor has $100,000 today and needs $106,000 at the end of one year, with no deposits or withdrawals in between. The implied return floor is $106,000 / $100,000 – 1 = 6%. If the portfolio earns 4%, it ends at $100,000 × 1.04 = $104,000. The investor has gained $4,000 but missed the goal by $2,000. All amounts and returns in this example are hypothetical.
Setting a floor identifies the event to measure. It does not establish that the floor is attainable at an acceptable level of risk. A deadline with intermediate withdrawals also needs more analysis than a single terminal-return comparison.
The Safety-First Ratio Formula
For a portfolio with positive return standard deviation, calculate:
- E(Rp) is the portfolio’s expected return for the chosen horizon.
- RL is the minimum acceptable return for that same horizon.
- σp is the standard deviation of the portfolio’s returns over that horizon and must be greater than zero.
The numerator measures the expected return cushion above the floor. Dividing by standard deviation expresses that cushion in units of volatility. Under normally distributed returns, the portfolio with the highest safety-first ratio has the lowest probability of missing the common floor.[1]
Keep the horizon, return definition and units consistent. For example, (12% – 6%) / 25% = 0.24 is equivalent to (0.12 – 0.06) / 0.25 = 0.24. Here, 25% standard deviation means 25 percentage points of return dispersion. Mixing 12 and 6 with a denominator of 0.25 gives the wrong scale. For input estimation, see standard deviation in finance.
If standard deviation is zero, the ratio is undefined. A truly certain return must instead be compared directly with the floor. Do not divide by zero or treat a rounded volatility estimate as proof that an investment’s return is certain.
From the Ratio to Shortfall Probability
The probability conversion requires an additional assumption: portfolio returns follow the stated normal distribution. Standardizing the floor subtracts the mean and divides by standard deviation.[2][3] That reverses the safety-first numerator:
Let Φ denote the standard normal cumulative distribution function: the probability to the left of a z-score. Then:
As SF rises, -SF moves farther left and the modeled shortfall probability falls. The ratio is a standardized distance, rather than a percentage probability. The following values illustrate the mapping; they are not universal investment-quality thresholds.
| Safety-first ratio | Floor relative to mean | Probability below floor |
|---|---|---|
| -1.00 | One standard deviation above | 84.13% |
| 0.00 | Equal to the mean | 50.00% |
| 1.00 | One standard deviation below | 15.87% |
| 2.00 | Two standard deviations below | 2.28% |
A negative ratio therefore means the floor exceeds expected return. Under the normal model, missing that floor is more likely than meeting it. A zero ratio gives a 50% modeled miss probability; it does not mean zero risk. For the underlying tail-area calculations and table conventions, see the normal distribution in finance.
Example: Choose Among Three Portfolios
Consider three hypothetical portfolios over the same one-year horizon. Their expected simple total returns are 8%, 10% and 12%, with respective return standard deviations of 12, 18 and 25 percentage points. Assume each return distribution is normal, use a common 6% floor, and ignore taxes, fees and intervening cash flows. These are invented inputs, not forecasts for actual funds.
| Portfolio | Expected return | Standard deviation | Safety-first calculation | Normal shortfall probability |
|---|---|---|---|---|
| A | 8% | 12 percentage points | (8 – 6) / 12 = 0.167 | 43.38% |
| B | 10% | 18 percentage points | (10 – 6) / 18 = 0.222 | 41.21% |
| C | 12% | 25 percentage points | (12 – 6) / 25 = 0.240 | 40.52% |
C ranks first because 0.240 is the largest safety-first ratio. Its larger expected cushion more than compensates for its higher standard deviation in this particular ratio. The probabilities use the unrounded ratios; displayed ratios are rounded to three decimals.
1. Find the cushion: 12% – 6% = 6 percentage points.
2. Divide by volatility: SF = 6 / 25 = 0.24.
3. Reverse the sign for the floor: zL = (6 – 12) / 25 = -0.24.
4. Evaluate the left tail: Φ(-0.24) = 0.405165…, or 40.52%.
C’s modeled miss probability is approximately 2.87 percentage points below A’s, calculated from the unrounded probabilities. This is a ranking among three alternatives. A 40.52% chance of missing the floor can still be unacceptable for the investor’s purpose.
What If the Floor Is 0% Instead?
Keep the same portfolio assumptions and change only the floor to 0%. The safety-first ratios become 8 / 12 = 0.667 for A, 10 / 18 = 0.556 for B and 12 / 25 = 0.480 for C, rounded. Their normal-model probabilities of a negative return are 25.25%, 28.93% and 31.56%, respectively.
A now ranks first. C was the preferred choice for avoiding a return below 6%; A is preferred for avoiding a return below 0% under these assumptions. Choosing the floor is part of specifying the investment problem, so document it before looking for a winning portfolio.
Roy’s Criterion vs. Sharpe and Sortino
All three ratios compare return with a risk measure, but their benchmarks and denominators answer different questions. The table uses an expected-return convention throughout.
| Measure | Numerator | Denominator | Interpretation |
|---|---|---|---|
| Safety-first | Expected return minus chosen floor | Total return standard deviation | Standardized cushion; maps to floor-miss probability under normality |
| Sharpe | Expected return minus benchmark return | Standard deviation of differential returns | Expected differential return per unit of its volatility |
| Sortino | Expected return minus chosen target | Downside deviation relative to that target | Target-relative return per unit of downside deviation |
For the usual Sharpe comparison with a constant risk-free return, subtracting that constant leaves standard deviation unchanged.[4] Suppose that benchmark is a hypothetical 2% over the same year. The three portfolios have Sharpe ratios of (8 – 2) / 12 = 0.500, (10 – 2) / 18 = 0.444 and (12 – 2) / 25 = 0.400, rounded. Sharpe ranks A first, while the 6% safety-first floor ranks C first.
The benchmark changed, which explains the different ranking. If the safety-first floor equals the same constant risk-free return and all other inputs match, its formula equals that Sharpe formula. A varying benchmark requires the standard deviation of differential returns; portfolio standard deviation alone is insufficient. Our Sharpe ratio guide covers the broader performance measure.
The Sortino convention shown here uses the same selected target in its numerator and downside deviation. Unlike Roy’s denominator, downside deviation focuses on below-target outcomes. Published calculations vary, so check the target, horizon and denominator method before comparing Sortino values.[5] Neither a Sortino value nor a Sharpe value is itself a shortfall probability.
How to Apply the Safety-First Criterion
- Define the event. Write the minimum acceptable return and the date when it matters. Distinguish a nominal principal floor from a spending, inflation-adjusted or other target.
- Make inputs comparable. Use the same horizon, currency, simple or log return definition, and fee/tax basis for every portfolio and its floor. A stated annual return target needs annual-horizon return estimates.
- Identify how the estimates were obtained. Label historical estimates versus forecasts and their sample or model. The ratio does not estimate expected returns or validate volatility inputs.
- Calculate the ratio. Subtract the common floor from each expected return, divide by its positive standard deviation, and retain unrounded results for ranking.
- Evaluate the distribution assumption. If using normal returns, calculate Φ(-SF). Otherwise estimate the target-miss probability using an appropriate explicitly specified model or scenario distribution rather than assuming the normal mapping.
- Test the decision. Compare the winning probability with the risk the goal can tolerate. Examine losses below the floor, liquidity needs and sensitivity to changed inputs before deciding whether any alternative is acceptable.
If every choice has a high modeled chance of shortfall, the useful conclusion may be to reconsider the goal, funding, horizon or available portfolios. Selecting the largest ratio alone does not solve a funding gap.
Common Mistakes
- Minimizing the ratio. Maximize SF; minimize the modeled shortfall probability. Their directions are opposite because the CDF input is -SF.
- Calling every shortfall a loss. A positive return can miss a positive floor. State exactly which event the reported probability describes.
- Reading 0.24 as a 24% probability. It is a standardized cushion. In the normal example, the corresponding miss probability is 40.52%.
- Changing floors across portfolios. That compares different goal events. Hold the floor and horizon fixed when selecting among alternatives.
- Mixing annual targets with monthly volatility. Obtain consistent horizon inputs; a convenient scaling shortcut needs assumptions of its own.
- Assuming a larger ratio always means less overall risk. An above-mean floor gives a negative ratio. With expected return fixed below that floor, more volatility can make the ratio less negative and improve the modeled chance of crossing the target, while also widening possible losses.
Limitations of Roy’s Criterion
The normal-CDF probabilities and their ratio ranking depend on the stated distribution and inputs. Means and standard deviations alone do not determine shortfall probabilities for arbitrary return distributions.
Return shape matters. Financial-return research documents heavy tails and persistent, changing volatility.[6] An unsuitable normal approximation can misstate probabilities or portfolio rankings. The normal distribution also extends without bound in both directions.[2] A fully paid, unlevered long-stock position cannot lose more than its initial value through the stock price alone, so a normal model of its simple return is an approximation.
Frequency does not measure severity. The criterion counts whether the floor is missed. It does not distinguish a tiny miss from a catastrophic one within that event. Assess the size of adverse outcomes separately. The probability of shortfall is also different from the tail-loss measure commonly called Expected Shortfall.
One terminal return does not describe the path. A portfolio may recover by the deadline but suffer a large interim drawdown, or be difficult to sell when cash is needed. Intermediate withdrawals, changing liabilities and trading constraints need their own analysis.
Estimated advantages can be fragile. Expected returns, volatility and distribution assumptions are uncertain. Small ranking differences do not establish dependable future superiority. Recalculate under plausible alternative assumptions and retain the actual target-miss probabilities alongside the ratios.
Frequently Asked Questions
Sources and Calculation Notes
Sources checked November 16, 2026. Portfolio inputs, dollar amounts and the 2% benchmark are hypothetical. Probabilities use the assumed normal model and unrounded inputs, with final percentages rounded to two decimals. Under a continuous normal distribution, the probability at the exact floor is zero, so using “below” or “at or below” gives the same model result.
- CFA Institute: Members’ Guide to 2022 Refresher Readings, Common Probability Distributions, printed pp. 97 and 101: safety-first learning outcome and summary. This public edition supplies mathematical background, not a claim about current exam coverage.
- NIST/SEMATECH: Normal Distribution, parameters, cumulative distribution and unbounded support.
- NIST/SEMATECH: Standard Normal Cumulative Distribution Function, standardization, table convention and tail probabilities.
- William F. Sharpe: The Sharpe Ratio, Journal of Portfolio Management, Fall 1994; differential-return definition and risk-free benchmark.
- Deborah Kidd: The Sortino Ratio: Is Downside Risk the Only Risk that Matters?, CFA Institute, 2012; target-relative calculation and convention caveats.
- Engle and Patton: What good is a volatility model?, Quantitative Finance 1 (2001), sections 2.1 and 2.5: volatility persistence and tails.
Disclaimer
This article is general educational information and does not constitute personalized investment advice. Hypothetical calculations depend on their assumptions and do not predict returns, guarantee a goal or protect principal. Evaluate your circumstances and the risks of any investment before acting.
