Variable Position Sizing: Conviction-Based Risk

Variable position sizing is justified only when a predeclared setup tier predicts a materially different return distribution—and that separation survives honest testing. “Conviction” must mean a frozen, observable classification made before the outcome, not how strongly a trader feels about the next trade.

If the tiers do not separate out of sample, use fixed risk. If they do, allocate within account and portfolio caps, include costs and dependence, and keep the classification record immutable.

Conviction Is a Model Output, Not a Feeling

Fixed sizing assumes each accepted signal receives the same risk allocation. Variable sizing allows allocation to change with evidence available before entry. The crucial word is before: if a grade changes after seeing the result, the history cannot test the rule used in real time.

A defensible conviction tier has:

• a named setup and version;
• observable inputs available at the decision timestamp;
• mutually exclusive classification rules;
• a documented treatment for missing or ambiguous inputs;
• a risk weight chosen independently of recent wins, losses, or urgency.

The trade-quality scoring guide shows how to keep process grade separate from outcome. A high-quality loss remains a high-quality loss if the pre-trade evidence was recorded correctly.

When Variable Sizing Earns the Right to Be Used

A higher historical win rate alone is insufficient. Compare average win and loss, costs, tail losses, concentration, sample uncertainty, and how often the tier was available. A tier driven by one market month or one exceptional trade is not a robust allocation signal.

Build the Tier Model Without Outcome Leakage

1. Freeze the base setup. Variable size should not hide several different entry systems under one name.
2. Choose candidate inputs. Examples include setup subtype, predeclared regime, liquidity state, session, volatility bucket, and verified context alignment. Use inputs the strategy can observe consistently.
3. Define the classifier. Write exact conditions for Tier A, B, C, and skip—or use fewer tiers. Do not backfill subjective labels.
4. Label chronologically. Apply the rules to records without seeing later outcomes. Preserve rejected and ambiguous cases.
5. Separate development from evaluation. Form the tiers on one period; measure the frozen version on untouched or forward evidence using the strategy testing pipeline.
6. Version every change. A new input, threshold, or missing-data rule creates a new tier model.

Confluence count, multi-timeframe alignment, and session timing can be candidate inputs, but they are not universal edge levers. Each must earn its place in the trader’s own frozen test.

Convert a Tier Into an Executable Size

First set the permitted account risk for the trade under the current daily, strategy, and portfolio budget. Then apply the tier weight and divide by estimated risk per tradable unit:

Estimated risk per unit includes entry-to-invalidation distance, tick or point value, fees and commissions, currency conversion, and a defensible allowance for slippage or gaps. Quantity must respect lot/contract increments, leverage, margin, and the platform’s order behavior.

The risk-per-trade framework supplies the account-level boundary. Variable sizing changes allocation inside that boundary; it does not authorize a larger daily or maximum-loss budget.

Choose Weights Conservatively

A simple implementation may assign the baseline tier a weight of 1, weaker-but-still-eligible evidence a lower weight, stronger validated evidence a modestly higher weight, and rejected evidence zero. Those are relative labels, not default percentages of account equity.

Simulate the complete sequence with the proposed weights. Compare net return, maximum and rolling drawdown, loss clusters, concentration, risk of account-limit breach, and sensitivity to worse fills. Test whether the apparent benefit survives smaller differences between tiers and shuffled or resampled outcomes under assumptions appropriate to the data.

Kelly’s original paper links an optimal fraction to known probabilities and odds in repeated bets. Trading estimates are uncertain, costs and dependence matter, and the distribution can change. The paper is mathematical context—not permission to insert an estimated win rate into a formula and risk the result. See Kelly’s original 1956 article.

Account and Portfolio Constraints Override the Tier

Several simultaneous “high-conviction” signals may be one economic bet. Aggregate exposure by underlying, direction, currency, sector, strategy, and event. Apply a portfolio cap before the individual tier map; otherwise variable sizing can concentrate risk exactly when signals are most correlated.

On a prop or restricted account, replay the weighted sequence through the exact program, phase, region, account size, daily/max-loss calculation, reset boundary, floating P&L treatment, consistency or concentration terms, holding restrictions, and platform behavior. The prop-firm drawdown guide explains why headline account size is not the same as usable risk capacity.

A tier does not override a hard account rule. When the remaining permissible loss is below one executable unit, size is zero.

Five Failure Modes to Audit

1. Grade inflation: more trades migrate upward without a versioned rule change.
2. Outcome editing: winners are relabelled higher and losers lower after the fact.
3. Drawdown pressing: risk increases to recover losses rather than because the frozen tier changed.
4. Tier fishing: inputs or thresholds are selected because they split historical results attractively.
5. Portfolio blindness: several individually valid sizes create one oversized correlated position.

Audit tier frequency, missing classifications, overrides, realized risk versus planned risk, and the proportion of total risk allocated to each tier. A rising top-tier share can mean better selection—or a drifting classifier. The immutable pre-trade record tells the difference.

When Fixed Sizing Is Better

Prefer fixed risk when tier definitions are subjective, the sample is too thin, out-of-sample separation disappears, setup quality is intentionally uniform, classification latency breaks execution, or account limits make the larger weight impractical. Systematic strategies may still vary size, but the sizing rule belongs in the tested system rather than a discretionary post-signal score.

Fixed sizing is not a beginner penalty. It is the correct baseline until variable allocation demonstrates incremental value after costs and within the same risk constraints.

Compare the candidate map with a fair control: the same eligible trades, timestamps, fills, exits, costs, and total risk policy under fixed sizing. If the variable version appears better only because it silently skips losing records, uses later information to grade them, or deploys more aggregate risk, the comparison is invalid. Report absolute and risk-adjusted outcomes alongside drawdown and concentration.

Keep operational complexity in the comparison too. A theoretically better map can be worse in practice when grades are frequently missing, quantities cannot express small differences, or latency changes fills. The decision is not “variable is sophisticated”; it is whether the added rule improves a reproducible, executable portfolio after every constraint is applied.

How TSB Keeps Conviction From Drifting

Trader’s Second Brain can save the setup and version, pre-trade tier inputs, assigned grade, planned risk, quantity, account, timestamp, costs, and actual outcome in one journal record. Reports and Backtester filters can compare frozen tiers, while the raw rows preserve losses, missing grades, and overrides.

Coach can ask whether tier frequencies or outcomes differ in a selected evidence set and surface gaps in tagging. That is the right strength: turning a large journal into inspectable questions. It should qualify or refuse the conclusion when tier rules changed, samples are sparse, or the user asks it to infer confidence or edge from fields that were never recorded.

TSB recognizes 330 exact import profiles and has normalized 600K+ imported trades. These figures describe import coverage and imported trade volume—not users, a conviction-tier sample, proof of edge, or promised returns.

TSB is our product. We disclose that ownership because this guide recommends its journal, Backtester, and Coach workflow.

Methodology Note

• Protected intent: the variable-position and conviction-based sizing job remains, but conviction is defined as a reproducible pre-trade model output.
• Research boundary: Kelly is cited only for its original probability/odds context; no retail allocation is derived from it.
• Removed claims: universal tier percentages, win-rate/R-multiple gaps, P&L lifts, sample thresholds, experience rules, audit cadences, and deterministic failure timelines were not retained.
• Product facts: journal fields, Backtester/Coach boundaries, and canonical import figures were checked against current server code.

For our evidence and correction process, see the editorial methodology.

Final Verdict: Size the Evidence, Not the Emotion

Variable sizing can improve allocation only after the tier itself earns trust. Freeze the classifier, validate it away from development data, weight it conservatively, and simulate the entire account path with costs and dependence.

When the separation is uncertain, fixed risk is the stronger decision. When a hard account or portfolio limit binds, the correct conviction-based size is zero.