A New Interpretation of Information Rate (1956)

Kelly connects information quality to the maximum long-run logarithmic growth of capital and shows why maximizing expected terminal wealth can imply ruinous all-in betting.

paper A New Interpretation of Information Rate Kelly, John L. Jr. · 1956 originalarchived pdf

This ten-page Bell System Technical Journal paper is the primary source behind the Kelly criterion. It studies a gambler who receives information through a noisy channel and can repeatedly reinvest capital. The archived PDF was read page by page and has SHA-256 d422b93805694672f64a1daacbc7396504d4718f55d115c45cd29afbc5a477bb.

Durable contribution #

Kelly's central warning is that maximizing expected capital after each bet can recommend staking everything. When the signal is fallible, this policy eventually loses all capital with probability one even though its expected terminal value can grow. Maximizing expected logarithmic growth instead balances growth against multiplicative ruin.

In the paper's even-money binary example, if a signal is correct with probability qq and wrong with probability 1−q1-q, the optimal signed fraction is the difference between those probabilities. That formula belongs to its exact payoff convention. It must not be pasted unchanged into a prediction-market contract bought at arbitrary price pp.

For a YES contract bought at pp, if ff is the fraction of bankroll spent on contracts and the trader's probability is qq, the wealth multipliers are

WYES=1+f1−pp,WNO=1−f.W_{YES}=1+f\frac{1-p}{p}, \qquad W_{NO}=1-f.

Maximizing

g(f)=qlog⁡WYES+(1−q)log⁡WNOg(f)=q\log W_{YES}+(1-q)\log W_{NO}

gives

f∗=q−p1−pf^*=\frac{q-p}{1-p}

when q>pq>p, before costs and under the stated independent, repeated, correctly specified model. If q≤pq\le p and shorting is excluded, the optimum for buying YES is zero.

Complete coverage #

  • Pages 1–2 introduce the communication-channel betting interpretation.
  • Pages 3–4 contrast expected capital with almost-sure logarithmic growth and solve the binary fair-odds case.
  • Pages 5–7 generalize to multiple symbols and arbitrary fair odds, linking maximum growth to information rate.
  • Pages 7–9 introduce a track take, capital held back, and the resulting constrained allocation.
  • Page 10 states limitations and the dominance claim only over fixed allocation strategies, then gives references.

Source claims audit #

Claim Verdict GTK treatment
All-in betting can maximize expected wealth while causing almost-sure eventual ruin current under the repeated model Use to separate arithmetic EV from survival and growth
Expected log growth gives the optimal fixed repeated allocation under the paper's assumptions current State assumptions and payoff convention every time
The even-money fraction can be used unchanged for a contract at any price false Re-derive from the contract's actual wealth multipliers
Full Kelly is robust to probability-estimation error and dependence false Treat model error and correlation as reasons to reduce aggregate exposure
Kelly proves dominance over every adaptive strategy not established in the paper Keep the source's explicit fixed-allocation boundary

Visual audit #

Every page was rendered and inspected. The paper contains no plots, diagrams, photographs, or information-bearing tables. Its displayed equations are the load-bearing visuals; they are represented as native mathematics in GTK rather than raster excerpts. Notation and inequality direction were checked against the page images because OCR alone is not authoritative for formulas.

Editorial role #

GTK uses Kelly as a position-sizing framework, not a promise of profit. The trading chapter derives the formula for a 0/10/1 contract, recommends explicit limits and fractional use when qq is uncertain, and treats correlated positions as one aggregate bet rather than independent Kelly opportunities.

Searches titles, aliases, headings, summaries, and full text.↑↓ move ↵ open esc close