5.7.5 Failure Modes & Practice

Basic Prerequisites: 04 · Fractional Kelly & Ruin and Statistics & Inference.

---

1. Intuition & Practical Objective

Every Kelly formula in the earlier pages takes $p,m,s$ as known. In real trading they are estimated from finite, noisy data - and the pivot of the whole folder is that $f^{\ast }$ is maximally sensitive to exactly the parameters that are hardest to estimate. The objective of this page is to make the failure modes concrete and to hand you the professional de-rating playbook.

Four failure modes dominate:

1. Parameter-estimation error → overbetting. If your estimated edge is too optimistic (a $+1\sigma$ draw on $p$) your "$f^{\ast }$" is an overbet, and an overbet of $2\times$ is already at the critical fraction - a turned-negative growth rate and near-certain ruin despite a real edge.
2. Fat tails. $g\approx f\mu -\frac{1}{2}{\sigma }^2{f}^2$ understates how bad the bad state is. A heavy loss-tail makes the exact log-optimal $f^{\ast }$ smaller than the Gaussian approximation, so sizing on the Gaussian number overbets.
3. Non-ergodicity / finite horizon. Kelly’s dominance is a long-run, almost-sure result; a single path with a finite horizon and a big early drawdown can kill a strategy that is asymptotically optimal.
4. Drawdown / leverage interaction. In the continuous levered case, overbetting near $f_c$ meets margin calls at exactly the worst moment - the drawdown and funding failures compound.

Takeaway. The single most important practitioner law in the whole folder: if you are not sure about your parameters, you must not bet full Kelly. De-rate to half Kelly as a default and further on parameter uncertainty - because overbetting is punished asymmetrically harder than underbetting, being wrong about $f^{\ast }$ in the upward direction is far more dangerous than being wrong downward.

---

2. Mathematical Ground Truth & Derivations

2.1 Estimation error: the overbet threshold

With a true win probability $p_{\text{true}}$ and an estimate $\hat{p}$ from $n$ trials, $\text{SD}(\hat{p})=\sqrt{p_{\text{true}}(1-p_{\text{true}})/n}$. If you bet full Kelly on $\hat{p}$ while the truth is $p_{\text{true}}$, your deployed fraction is $f=2\hat{p}-1$. The true growth rate at that $f$ is

$$g_{\text{true}}(f)=p_{\text{true}}\ln (1+f)+(1-p_{\text{true}})\ln (1-f).$$

If $\hat{p}>p_{\text{true}}$ pushes $f$ above the true critical $f_c^{\text{true}}$, $g_{\text{true}}(f)<0$ and the strategy turns into a guaranteed-loss machine - the estimate flipped the sign of the realised growth rate.

Concretely (below): truth $p=0.52$, $\hat{p}=0.57$ (a $+1\sigma$ estimate from $n=100$), so you bet $f=0.14$; true Kelly is $0.04$ and the true $g(0.14)=-0.0043<0$. A 3.5× overbet. The realised outcome: median wealth collapses to ~0 and ~73% of accounts are under 1% of starting capital in 2,000 bets.

2.2 Fat tails understate the true ruin

The Gaussian/continuous approximation $g\approx f\mu -\frac{1}{2}{\sigma }^2{f}^2$ is exact only for log-normal returns. A heavy-loss tail (jump risk) contributes large negative log-returns that the second-order expansion misses. The exact log-optimal $f^{\ast }=\arg {\max }_f{\sum }_i{p}_i\ln (1+f{r}_i)$ is then smaller than the Gaussian $\mu /{\sigma }^2$ estimate, so a Gaussian-sized bet is an overbet in tail reality. Worked below: exact $f^{\ast }=0.581$ vs Gaussian $f_{\text{gauss}}^{\ast }=0.766$, and the Gaussian bet materially raises the ruin probability.

---

3. Computational Implementation - demonstrated failure modes

Stdlib only. (a) Estimation error flipping a positive edge into ruin; (b) heavy-tail log-optimal vs Gaussian sizing.

The two blocks are the folder’s two most important “do not do this” results. (a) A single-standard-deviation optimism about $p$ turns your positive-edge strategy into one that loses 99.995% of capital with 72.8% probability of ending below 1% - while the correctly-sized true-Kelly bet grows a bankroll $4.9\times$ with zero such ruin. (b) Even with an exact mean and variance, the Gaussian Kelly $f^{\ast }=0.766$ is 32% larger than the true log-optimal $f^{\ast }=0.581$ because it ignores the $-60\%$ tail; sizing on the Gaussian number is a silent overbet (an overbet of ~32%, which steepens the ruin curve sharply).

---

4. Failure Modes & First-Principles Breakdowns

1. Overbetting is asymmetric in danger. An underbet costs a little growth; an overbet past $f_c$ flips the sign of the growth rate. If you must err, err down - hence half Kelly as the operating default. This is the single sentence to carry out of this folder.
2. Gaussian variance is not SD of log growth under jumps. Fat tails mean the true drag exceeds $\frac{1}{2}{\sigma }^2$; jump risk is invisible to a second-order expansion. Estimate the exact $\mathbb{E}[\ln (1+fR)]$ over the empirical/parametric heavy-tailed distribution, not a normal approximation.
3. Backtest optimism. $p,m,s$ measured in-sample over a regime shift are systematically optimistic; the “edge” a backtest reports is a biased input to $f^{\ast }$. Fractional Kelly here is model-risk management, not timidity.
4. Leverage × drawdown × margin. In the continuous case an overbet near $f_c$ produces drawdowns that trigger margin calls at the trough; the forced de-lever locks in ruin. De-risking must happen pre drawdown, which fractional Kelly does by construction.

---

5. References

• Thorp, Edward O.: The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market (2006)
• MacLean, Thorp & Ziemba (eds.): The Kelly Capital Growth Investment Criterion (2011)
• MacLean, Ziemba & Blazenko: Growth versus Security in Dynamic Investment Analysis, Management Science (1992)
• Chopra & Ziemba: The Effect of Errors in Means, Variances, and Covariances on Optimal Portfolio Choice (JPM 1993)
• Bouchaud & Potters: Theory of Financial Risk and Derivative Pricing

---

6. Connected Graph Bridges

• Back: 04 · Fractional Kelly & Ruin · Index Hub
• Estimation-error sibling: Covariance Shrinkage & RMT · 05 · MPT Failure Modes
• Fat tails: Extreme Value Theory & Fat Tails · 06 · Advanced Extensions (foundations)
• Continue: 06 · Advanced Extensions