> ## Content Index
> Fetch the complete content index at: https://aligrithm.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# 3.48 The Information-Survival Ratio: Diagnosing Conditioning Overfit
- URL: https://aligrithm.com/the-information-survival-ratio-diagnosing-conditioning-overfit/
- Published: 2026-11-16T09:57:00.000Z
- Updated: 2026-11-16T09:56:59.000Z
- Description: Walk-forward conditioning overfit: divide the selected Kullback-Leibler budget by the in-sample budget. That ratio is the information-survival ratio of one strategy.
- Author: ali askar
- Tags: 3. Robust Systems Lab

Rodríguez Domínguez lets a BarclayHedge style index spend a Kullback-Leibler budget, in nats, on a state-dependent reweight of its own return, then asks how much of that budget blocked cross-validation still selects. The information-survival ratio is the walk-forward budget divided by the in-sample budget. A walk-forward budget of 0.10 nats against an in-sample budget of 0.20 nats is a ratio of 0.50\. The deflated Sharpe ratio, and the probability of backtest overfit in the old article ["CSCV: A Direct Probability of Backtest Overfit"](https://aligrithm.com/cscv-a-direct-probability-of-backtest-overfit/), score the search across strategies. The information-survival ratio scores the state variables inside one strategy.

## The kernel is an exponential tilt, and the budget is its divergence

You reweight each period's excess return by a positive number whose average is 1\. The number can depend on a state known before the period opened. Rodríguez Domínguez calls that number a conditioning kernel. A kernel that is identically 1 is the unconditional book: you hold the return as published.

The kernel he fits is an Esscher tilt on a feature map you fix in advance. The feature map, phi, turns the state into a short vector. The tilt, theta, is a vector of the same length. The kernel is the exponential of theta times phi, divided by the average of that exponential, so the weights still average to 1.

$$ m\_\\theta(X) = \\frac{\\exp(\\theta^\\top \\phi(X))}{\\mathbb{E}\[\\exp(\\theta^\\top \\phi(X))\]} $$

The numerator levers periods whose features point the same way as theta. The denominator is the average of that lever, so a period with an ordinary feature gets a weight near 1, and the weights across periods average to 1\. Theta at zero gives a weight of 1 in every period.

The budget is the Kullback-Leibler divergence of this kernel from the flat weight, in nats per period. A nat is the natural-log unit of that divergence. The budget is the average of weight times log weight, and it equals zero when every weight is 1.

$$ \\kappa(m) = \\mathbb{E}\[m(X) \\log m(X)\] $$

On one standard normal feature, the kernel in his simulation remark is the exponential of theta times the feature, minus theta squared over 2\. He also prints the tilted moment: the average of weight times the feature equals theta. The weights average to 1 by construction. The log of this kernel is theta times the feature, minus theta squared over 2\. Average weight times that log. You get theta squared, minus half of theta squared, which leaves theta squared over 2.

$$ \\kappa = \\frac{\\theta^2}{2} $$

That identity is the arithmetic on the kernel and the moment he prints. He does not label it as a numbered equation. A budget of 0.20 nats on this one-feature kernel needs a tilt of the square root of 0.40, which is 0.632\. The square 0.632 times 0.632 equals 0.399424, and half of that is 0.199712, which is 0.200 nats to three decimals. A tilt of 0 spends 0 nats, and the book is the raw return. The empirical grid runs from 0 to 0.20 in steps of 0.02, eleven points. On this one-feature Gaussian kernel the top of the grid is the 0.632 tilt. The four feature sets in the tables have several columns, and there the budget stays the average of weight times log weight.

![](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_491-visual-0.png)

## Clipped mean-variance is the score the grid maximises

The object you maximise at a fixed budget is a mean-variance score of the reweighted return. Lambda is the charge on variance. The tables use lambda equal to 0.5\. The Sharpe ratio of the same reweighted return is a number you report after the budget is chosen.

$$ \\mathcal{V}(m) = \\mathbb{E}\[m r\] - \\lambda \\, \\mathrm{Var}(m r) $$

The first term is the average reweighted excess return. The second term charges lambda times the variance of that reweighted return. At lambda 0.5, a mean of 0.01 and a variance of 0.02 leave the score at zero, because 0.5 times 0.02 equals 0.01\. One unit of mean buys two units of variance. That is the exchange he describes. On the equity portfolios, 38 of 44 names keep the same selected budget at lambda 0.25, 0.5, and 1.0, and the other 6 move by one grid step of 0.02\. The asset-class mean ratio moves by less than 0.05 across that range. That stability is a fact about this panel. A lambda near zero, or a lambda large enough to force a corner, is outside the check.

The sample score uses winsorised returns. M is the clip, set to three in-sample standard deviations.

$$ r^{(M)} = \\max(-M, \\min(r, M)) $$

A month outside the band enters the score at the band. Suppose the in-sample standard deviation is 0.04, so M is 0.12\. A return of -0.20 enters as -0.12\. The 0.08 you cut off never reaches the mean or the variance that pick the budget. This 0.04 is a sketch of the clip, chosen so the band is easy to see. The clip puts the score inside a finite band, because the kernel is bounded once phi is bounded, and the clipped return is bounded by M. The band is what lets a Bernstein inequality deliver an exponential miss rate for the budget selector. Unclipped returns would leave a polynomial Markov bound.

The fitted score on a block divides by the count of months in the block.

$$ \\widehat{\\mathcal{V}} = \\bar{y} - \\lambda \\left( \\overline{y^2} - \\bar{y}^2 \\right), \\qquad y\_t = m\_\\theta(X\_t)\\, r\_t^{(M)} $$

Bar y is the average of the reweighted clipped returns. Bar y-squared is the average of their squares. The term in parentheses is the variance with divisor equal to the block length. A divisor of block length minus one is a different estimator.

Two hypothetical months, returns 0.04 and 0\. Flat weights, both 1\. The reweighted returns are 0.04 and 0\. The mean is 0.02\. The mean of the squares is 0.0008\. The squared mean is 0.0004\. The variance is 0.0004\. At lambda 0.5 the penalty is 0.0002 and the score is 0.0198.

Weights 1.5 and 0.5 still average to 1\. The reweighted returns are 0.06 and 0\. The mean is 0.03\. The mean of the squares is 0.0018\. The squared mean is 0.0009\. The variance is 0.0009\. The penalty is 0.00045 and the score is 0.02955\. The mean rose by 0.01\. The penalty took back 0.00025\. The score rose by 0.00975, from 0.0198 to 0.02955\. These two months are a sketch of the functional, so the divisor is visible. Sum the squared deviations and divide by 1 instead of by 2, and the variance doubles.

He maximises this score subject to the sample divergence staying at or below the budget on trial. The Sharpe ratio has the standard deviation of the reweighted return in the denominator. A resample that makes that standard deviation tiny sends the ratio through the ceiling, and the argmax jumps. The clipped score cannot do that.

## The survival ratio divides the two caps

A walk-forward cap of 0.10 nats on an in-sample budget of 0.20 nats is 0.50, and the average that picks the cap is a four-fold score with a gap of unused months around each held-out block.

Split the sample into four contiguous folds. For each fold, hold out that block, drop a buffer of months on either side of it, fit theta at the candidate budget on what remains, and record the held-out score. The walk-forward score of a budget is the average of those four held-out scores. The walk-forward cap is the budget with the highest average. Ties go to the smaller grid point.

$$ \\widehat{\\mathrm{CV}}(\\kappa) = \\frac{1}{4} \\sum\_{j=1}^{4} \\widehat{\\mathcal{V}}\_{j}(\\hat{\\theta}\_j(\\kappa)), \\qquad \\widehat{\\kappa}\_{\\mathrm{wf}} = \\arg\\max\_{\\kappa \\in \\mathcal{K}} \\widehat{\\mathrm{CV}}(\\kappa) $$

Theta-hat sub j is the tilt fitted on fold j's training months at that budget. The grid is the eleven points from 0 to 0.20.

The buffer width in Theorem 8 is the floor of the sample length to the power one third.

$$ b\_T = \\lfloor T^{1/3} \\rfloor $$

On a hedge-fund index with T equal to 200, the minimum length he requires in that panel, the cube root is 5.85\. The floor is 5 and the ceiling is 6\. The stated training cost is two sides, times four folds, times the width. At the floor, 2 times 4 times 5 equals 40 months, a fifth of the sample, excluded from training. At the ceiling, 2 times 4 times 6 equals 48 months. Theorem 8 and the setup section print the floor. The introduction and the caption on the joint specification bootstrap print the ceiling. The source disagrees with itself by one month at T = 200\. The formula above keeps the theorem's floor. One of the two expressions is wrong, or the convention changed and the other line stayed.

The ratio is the walk-forward cap divided by the in-sample budget.

$$ \\mathrm{ISR} = \\frac{\\widehat{\\kappa}\_{\\mathrm{wf}}}{\\kappa\_{\\mathrm{in}}} $$

A walk-forward cap of 0.10 against an in-sample budget of 0.20 gives 0.50\. The same walk-forward cap against 0.10 gives 1\. Against 0.40 it gives 0.25\. The tables fix the in-sample budget at 0.20 and stop the grid at 0.20, so every printed ratio sits between 0 and 1 by arithmetic. Set the in-sample budget to 0.40 and leave the grid topped at 0.20, and the largest ratio you can print is 0.20 divided by 0.40, which is 0.50, whatever the held-out months wanted. He sets the 0.40 check aside for that reason. Extend the grid past the budget you intend to quote. If the denominator is instead the budget the in-sample score itself picked, and the walk-forward grid is allowed to run past it, a walk-forward pick of 0.14 against an in-sample pick of 0.10 prints 1.4\. That 1.4 illustrates the warning. The tables use a fixed denominator of 0.20\. Restrict the walk-forward grid to budgets at or below the in-sample pick, and the ratio stays inside the unit interval.

The printed ratio records which ceiling the held-out folds voted for. The fitted kernel's realised divergence is a separate number. The optimiser may overshoot the cap by 10 percent as a numerical tolerance, and that slack is a solver setting. In fewer than 2 percent of the positive-budget fits, the constraint binds to within 1 percent. The selected grid point is usually a ceiling the tilt sat inside. A ratio of 1 says the held-out folds still wanted the most aggressive ceiling on offer. The nats the kernel consumed can be lower, and he reports that consumption in the replication files, off to the side of the ratio.

## The selector needs a gap between neighbouring caps

At 2,000 observations, a gap of 0.015 between the best budget and the next one produces a cross-fitted hit rate of 0.206\. The walk-forward cap is an argmax over eleven points, so its sampling noise is a selection problem. Write kappa-star for the budget that maximises the population cross-validated score, and write delta for the gap between that score and the next-best budget on the grid.

$$ \\Delta = \\mathrm{CV}(\\kappa^\*) - \\max\_{\\kappa \\neq \\kappa^\*} \\mathrm{CV}(\\kappa) $$

A delta of zero means two budgets are tied in population, and the sample pick stays random however long the series gets. Theorem 8 assumes delta is a fixed positive number. Under that assumption the infeasible selector, the one that plugs in the population tilt and only then picks the budget, misses kappa-star with probability at most (K minus 1) times an exponential in the sample length times delta squared.

$$ \\mathbb{P}(\\widehat{\\kappa}\_{\\mathrm{wf}}^{\\mathrm{inf}} \\neq \\kappa^\*) \\leq (K - 1) \\exp(-c \\, T \\, \\Delta^2) $$

K is 11, so K minus 1 is 10\. The constant c is positive and depends on the mixing rate, the kernel bounds, the clip, and the fold count. He does not give a numerical c. The display is a rate. The feasible selector, the one that refits theta inside each training block, adds a second exponential from the cross-fit error. That term decays with delta, because the cross-fit remainder is a product of two concentrated pieces and the inequality is applied at a deviation of delta over 4\. The miss probability of the selector you can run is then bounded by a constant times an exponential of minus T times the slower of the two rates.

$$ \\mathbb{P}(\\widehat{\\kappa}\_{\\mathrm{wf}} \\neq \\kappa^\*) \\leq C\_\\Delta \\exp(-c\_\\Delta T), \\qquad c\_\\Delta = \\min(c \\Delta^2,\\, c\_1 \\Delta) $$

C-delta and c-1 are positive constants he does not tabulate. A wide gap helps both terms. A narrow gap is limited by whichever decays more slowly.

Look at the two solid lines. The upper line is the edge regime: the population budget is 0.02 and the gap is 0.163\. The lower line is the interior regime: the population budget is 0.08 and the gap is 0.015.

![Selection probability of the true budget against sample size. Upper solid line: gap 0.163. Lower solid line: gap 0.015. Dashed lines are blocked cross-validation without the extra buffer](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_491-selection_probability.png)

The cross-fitted hit rate in the wide-gap regime rises from 0.288 at 200 observations to 0.735 at 2,000\. Dividing 0.735 by 0.288 gives 2.55, the fold increase he reports. In the narrow-gap regime the hit rate rises from 0.069 to 0.206 over the same lengths. At 2,000 observations the narrow-gap hit rate is 0.206 against the wide-gap 0.735, so 1 minus 0.206, which is 0.794, of the narrow-gap replications miss the true budget. The wide gap divided by the narrow gap is 0.163 divided by 0.015, which equals 10.8667, and that square is 118.08\. The exponent uses that square. The hit rate at 2,000 observations is the pair 0.735 and 0.206.

The standard blocked selector, with no buffer beyond the fold split, posts the higher hit rate in every cell of the table. The largest difference is 0.232 minus 0.206, which is 0.026\. He summarises the column as at most about 0.03\. On this simulation the two selectors land in nearly the same place, and the standard one is slightly ahead. The buffer's job here is the near-independence of the training fit and the held-out score, which the feasible-selector bound uses. A book whose labels overlap the training window is where that independence earns its keep. This data-generating process has little of that overlap, and the hit rates show it.

The gap has to come from the conditional mean. A linear conditional mean, return equal to a times the lagged feature plus noise, produces a reweighted mean of a times theta under the Gaussian kernel he writes. With the divergence equal to theta squared over 2, theta is the square root of 2 times kappa, and the reweighted mean is linear in the square root of kappa. Neighbouring budgets then sit too close for the gap to clear bootstrap noise at sample sizes of 1,000 and below. A quadratic term in the conditional mean bends the score.

$$ \\mathbb{E}\[m\_\\theta r\] = a \\theta - b \\theta^2 $$

This is the mean he derives from the tilted moments, for a return equal to a times the feature minus b times the feature squared, recentered. The variance penalty is extra. He computes the population budget from the full score, and he states that the grid optimum is different from a squared over 8 b squared, which is where the mean alone would peak.

In the edge calibration, a is 0.6 and b is 1\. The mean-only tilt is 0.3\. The mean-only budget is 0.045\. The reweighted mean at that tilt is 0.6 times 0.3 minus 0.09, which is 0.09\. The full mean-variance budget on the grid is 0.02\. At theta 0.2, the tilt that spends 0.02 nats, the reweighted mean is 0.6 times 0.2 minus 0.04, which is 0.08\. The penalty is what pulls the optimum from the mean-only 0.045 down to his 0.02\. I am not recomputing the variance term, so 0.02 stays his number. The interior calibration, a of 1.5 and b of 1, has a mean-only budget of 0.281, above a grid that stops at 0.20, and the full-score optimum he reports is 0.08\. Quote 0.08\. The mean-only 0.281 is a different object, and he says so.

A conditional mean close to linear is the usual case on a return series. The simulation puts in the quadratic so the theorem has a fixed gap to chew on. On the panels, the per-name pick is noisy, and the ordering he trusts is the average across names.

## Combined state holds on hedge-fund indices and fades on single names

Look at column D, the hedge-fund row against the single-name row. The colours are medians.

![Median information-survival ratio by asset class and conditioning set. Column D is the combined state: the hedge-fund cell prints 1.00 and the single-name cell prints 0.25](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_491-median_isr_heatmap.png)

The means in the table are 0.97 on the 26 hedge-fund indices and 0.31 on the 12 single names. Equity portfolios on realised volatility, column B, are the third mean, 0.80\. The colours in those three cells are 1.00, 0.25, and 1.00\. The hedge-fund combined interval is 0.91 to 1.00\. The single-name combined interval is 0.15 to 0.50\. The hedge-fund interval sits entirely above the single-name interval. The equity volatility interval is 0.69 to 0.91.

The denominator on every mean is the fixed budget 0.20\. A mean of 0.97 is an average walk-forward cap of 0.194 nats, because 0.97 times 0.20 equals 0.194 to the two decimals he printed. A mean of 0.31 is an average cap of 0.062\. A mean of 0.80 is an average cap of 0.160\. Each name's cap is one of the eleven grid points, and the average falls between points. The medians snap to that grid after dividing by 0.20, so a cell can print 0.00 in colour with a mean far from zero. Equity combined, column D, prints 0.00 and has a mean of 0.40\. Hedge-fund volatility prints 0.05 and has a mean of 0.43\. Screen on the mean. The colour is the middle name, and the middle name on an eleven-point grid is a coarse object.

Lagged own-return, last period's return and its square, has means 0.64, 0.53, 0.76, 0.67, 0.92, and 0.28 across equity portfolios, industries, US size-momentum, developed-international size-momentum, hedge funds, and single names, and the macro set (lagged log changes in the VIX, the dollar index, and a commodity index) finishes behind the panel leader in every row. Realised volatility is two lags of log realised variance: the within-month sum of squared daily returns when the daily series exists, and a squared rolling six-month standard deviation of monthly returns otherwise. The combined column is lagged return, one lag of log realised variance, and the lagged VIX change. All four sets were fixed before the ranking. Features are standardised on the estimation sample and clipped to the interval from -3 to 3, a different clip from the three-standard-deviation clip on the returns that enter the score. The equity panel is 44 Fama-French portfolios from July 1963, four factors plus forty deciles, with the SMB factor dropped by an unconditional-mean screen before the diagnostic runs. The single names are one US large cap per sector, daily from 2001 to 2023, restricted to continuous histories, so a delisted name never enters. Their 0.31 is a fraction of a per-day budget of 0.20 nats. The hedge-fund 0.97 is a fraction of a per-month budget of 0.20 nats. Compare the fractions. The nats sit on different clocks.

## Charge the spread, and read the ratio next to a trial count

Leave the desk with four numbers on one strategy: the walk-forward cap, the ratio, the walk-forward Sharpe ratio of the book you would trade, and a cross-strategy correction. The correction is a deflated Sharpe ratio, here the probability that the Sharpe clears a hurdle adjusted for 70 trials and for non-normality, or the probability of backtest overfitting from the old article ["CSCV: A Direct Probability of Backtest Overfit"](https://aligrithm.com/cscv-a-direct-probability-of-backtest-overfit/). He marks a strategy high on the deflated Sharpe when that probability is at least 0.90, and high on the ratio when the ratio is at least 0.50\. The 70 trials are the 44 equity portfolios plus the 26 hedge-fund indices, and he uses that same 70 on every panel. On the equity portfolios, 15 of 44 clear the deflated-Sharpe bar and print a ratio below 0.50\. That cell is a conditioning fit the trial correction misses. On the hedge-fund indices, 10 of 26 miss the deflated-Sharpe bar and print a ratio of at least 0.50\. Nine of the 12 single names sit in the cell where both are low. The cross-tabulation is one ratio per strategy. The table states the two thresholds and does not name which of the four conditioning sets supplied the ratio. If you count the four sets as extra trials, his own caveat is to rescale 70 up to 280 before you deflate. He prints the version at 70 because this ratio is already the within-strategy half of that search.

The old article "The Real IS→OOS Sharpe Haircut (Closed Form, Not "50%")" prices a different fraction, out-of-sample Sharpe over in-sample Sharpe, for a linear book. Use that fraction on the Sharpe. Use this ratio on the state. Each one leaves the other's question open.

The paper charges no spread. The kernel is a position, it moves with the state, and the tables do not report a turnover. I am against sizing a book from the ratio until that turnover has paid a spread, a fee, and, where you are short, a borrow. The old article ["The Signal Ceiling: Why No Single-Bar OHLCV Edge Beats Costs in MNQ"](https://aligrithm.com/the-signal-ceiling-why-no-single-bar-ohlcv-edge-beats-costs-in-mnq/) is the walk-forward that did charge costs. Single-bar OHLCV families on that contract did not get through.

I am against reading the hedge-fund row as a signal you can trade. Those indices are AR(1)-unsmoothed when a Ljung-Box test at lag 12 rejects, and after that treatment the panel's average lag-1 autocorrelation of returns is 0.23 and of absolute returns is 0.26\. A state can look durable because the reported series is smooth. The single-name row is the noisy case, on a continuous-history screen, which is a survivor screen.

One name's bootstrap interval is too wide to trade. Across the panels the average lower end sits below 0.10 and the average upper end sits near 1\. The panel mean is the object that resolves. Hedge-fund indices put average bootstrap mass of 0.92 on the modal cap. That is a concentrated panel vote, on a grid that stops at 0.20\. Halving the budget and the grid together, from 0.20 to 0.10, keeps the cross-asset ordering and moves the mean ratio by less than 0.05\. A budget above 0.20 was not on the grid. Seed-to-seed refits of the optimiser stay within one grid step, 0.02, for all but one strategy in the equity, hedge-fund, and single-name panels.

The empirical settings around the four folds are an initial window of one third of the sample and an annual refit, every 12 months on the monthly panels and every 252 trading days on the daily names. He lists those knobs with the fold count and gives no formula that nests the annual refit inside the four folds. Run the selector as the four-fold argmax, and keep the initial window and the refit cadence as the walk-forward settings he fixed.

If you also let the data pick the feature set, the joint bootstrap on equities puts mass 0.51 on lagged return and 0.49 on realised volatility. That split is a tie. The joint corollary needs a positive gap on the product of specifications and budgets, and a tie is a gap of zero. He publishes the four sets side by side. Choosing the column with the highest mean after seeing all four is a second search. The ratio, computed set by set, leaves that search uncorrected.

## KEY POINTS

- Compute the information-survival ratio as the walk-forward Kullback-Leibler cap divided by the in-sample cap. A cap of 0.10 nats on a budget of 0.20 nats is 0.50\. The tables fix the denominator at 0.20.
- On one standard normal feature the divergence is theta squared over 2, so the tilt that spends 0.20 nats is 0.632\. Select the cap by clipped mean-variance at lambda 0.5\. Report the Sharpe ratio after the cap is chosen.
- Separate training and held-out months by the floor of T to the one third. At T = 200 that width is 5 and it drops 40 training months. The introduction prints the ceiling, width 6, which drops 48\. The source disagrees by one month.
- Average across names. The combined-state mean ratio is 0.97 on 26 hedge-fund indices and 0.31 on 12 single names. Equity portfolios on realised volatility sit at 0.80\. The heatmap colours are medians, and equity combined prints 0.00 against a mean of 0.40.
- Put a deflated Sharpe, or the probability of backtest overfitting, next to the ratio. Fifteen of 44 equity portfolios clear a 70-trial deflated Sharpe and print a ratio below 0.50\. The paper charges no spread, and the kernel is a position.

## References

- [The Signal Ceiling: Why No Single-Bar OHLCV Edge Beats Costs in MNQ](https://aligrithm.com/the-signal-ceiling-why-no-single-bar-ohlcv-edge-beats-costs-in-mnq/)
- [The Real IS→OOS Sharpe Haircut (Closed Form, Not "50%")](https://aligrithm.ghost.io/the-real-is-oos-sharpe-haircut-closed-form-not-50/?ref=aligrithm.com)
- [CSCV: A Direct Probability of Backtest Overfit](https://aligrithm.com/cscv-a-direct-probability-of-backtest-overfit/)
- [The Information-Survival Ratio: A Within-Strategy Diagnostic of Conditioning Overfit in Walk-Forward Backtests, Alejandro Rodríguez Domínguez](https://papers.ssrn.com/sol3/papers.cfm?abstract%5Fid=6905139&ref=aligrithm.com)

---

**A note on AI.* The ideas, research, analysis, and conclusions in this article are my own. I use AI tools to help with editing and wordsmithing, because English is not my first language, and I am not shy about that. AI-generated ideas and AI-assisted writing are not the same thing: the first is empty slop from a generic prompt, the second is a tool for communicating years of real research more clearly. Judge the work by its substance, not by whether software helped polish the prose.*