Every prediction comes with uncertainty. A signal to buy Bitcoin means little without knowing how much trust the model itself places in that call. Position sizing is the bridge: it translates confidence into capital allocation.
Most traders skip the bridge entirely. They see a buy signal and allocate the same amount they always do, whether the model's confidence is 0.55 or 0.90. That guarantees low-quality signals receive the same capital as high-quality ones, which quietly caps how much a good model can do for you.
This guide covers the practical approaches, from a simple multiplier framework to fractional Kelly, along with the portfolio-level limits that keep any of it from going wrong.
The prerequisite: the score has to be calibrated
Everything below assumes one thing, and it is the thing most often missing: the confidence score must be calibrated. Calibrated means the number behaves like a probability. Across all the calls where the model says 0.8, about 80% should work out, misses included.
This is not pedantry. Every sizing formula in this post takes a probability as input. Feed it a vanity score, the kind that exists to make a dashboard look confident, and the formula will faithfully concentrate your capital wherever the marketing dial pointed. Sizing amplifies whatever the score really is: information or decoration.
If you have no reason to believe the score you use is calibrated, from the provider's stated methodology or from your own forward log, that is the problem to solve before this one. And keep expectations realistic about what verification means: live systems recalibrate continuously, so you are checking that the scores behave like probabilities over your window, not matching a frozen chart.
The core principle
Higher confidence, larger position. Lower confidence, smaller position, or no position at all.
It sounds obvious, and almost nobody does it systematically. Done consistently, it changes the shape of your returns: your capital concentrates in the calls where the model genuinely knows something, and your worst calls are also your smallest.
A simple framework: base size and multiplier
Start with a base position size, the amount you would allocate to a standard trade. A common choice is a fixed fraction of the portfolio, say 2%.
Then scale the base by confidence. One workable scheme, with a $10,000 portfolio and a 2% ($200) base:
| Confidence | Multiplier | Position size |
|---|---|---|
| 0.90+ | 1.5x | $300 |
| 0.80 to 0.89 | 1.0x | $200 |
| 0.70 to 0.79 | 0.75x | $150 |
| 0.60 to 0.69 | 0.5x | $100 |
| Below 0.60 | 0x | $0, no trade |
The exact multipliers are illustrative. What matters is the structure: monotonic in confidence, with a floor below which the answer is zero.
That zero row deserves emphasis. Not every prediction deserves capital, and a sizing system needs a level at which it refuses to trade. This is the same philosophy behind why Pearlixa signals can come back as HOLD: a weak setup is not a small opportunity. It is no opportunity wearing a small position.
Adding risk/reward
Confidence is not the only input. A 0.75-confidence prediction with a 3:1 reward-to-risk ratio is a different animal from a 0.75-confidence prediction at 1:1.
Expected value combines them:
EV = (p × average win) − ((1 − p) × average loss)
When predictions arrive with explicit target and stop-loss levels, the reward and risk of each trade are computable up front, so this stops being abstract. Some traders use a composite score, confidence multiplied by the reward-to-risk ratio, and allocate more to higher composites. The effect is to concentrate capital in trades that are both likely and asymmetric, which is where compounding actually comes from.
Note that EV is only computable because the confidence is a probability. This is the calibration prerequisite showing up again.
Kelly, and why you should stay well under it
The Kelly criterion is the classical answer to "how much, exactly?". For a trade with win probability p and a payoff of b units won per unit risked:
f* = p − (1 − p) / b
f* is the fraction of capital that maximizes long-run growth. Example: p = 0.60 and b = 2 gives f* = 0.60 − 0.40/2 = 0.40, or 40% of capital.
That number should alarm you, and the alarm is correct. Full Kelly assumes p and b are known exactly. In markets they are estimates, and estimation error in p is punished brutally: bet full Kelly on an overestimated edge and the formula that maximizes growth becomes a formula for drawdowns you cannot psychologically or financially survive.
The practical uses of Kelly are two, and neither is "bet f*":
- Fractional Kelly as an operating size. Practitioners commonly size at a quarter to a half
- Full Kelly as a hard ceiling. Whatever your sizing scheme says, if it ever suggests a
And once more: f* takes p as input. An uncalibrated confidence fed into Kelly does not produce optimal growth. It produces confident ruin.
Portfolio-level constraints
Individual position sizing has to fit inside portfolio-level limits, because correlated positions are one position in disguise. Five simultaneous 0.9-confidence longs on five crypto assets are not five independent bets; in a stress event they move as one.
The usual constraint set: a maximum exposure per asset, a maximum total exposure across all positions, a correlation limit that treats highly correlated positions as shared risk, and a cap on the number of simultaneous positions. When several high-confidence opportunities appear at once and the limits force each position smaller, that is the system working, not failing.
Rules for adjustment, set in advance
Positions evolve. A trade moves in your favor and the stop can trail. A new prediction arrives with lower confidence and part of the position can come off. Confidence rises on fresh information and adding may be justified, if the portfolio limits still hold.
The content of the rules matters less than their timing: they must exist before the trade is on. Sizing decisions made mid-trade, under the pressure of an open profit or loss, are reliably worse than the same decisions made in advance. The entire value of systematic sizing is that it removes the moment-of-decision from the moment-of-emotion.
Implementation
For automated traders, this whole pipeline is code: a prediction arrives via API with confidence, target, and stop attached, the system computes size from the framework and the portfolio state, and execution follows.
For manual traders, a spreadsheet is genuinely enough. Confidence and prices in, position size out, and the sizing decision is made by the sheet rather than by the mood.
Either way, consistency beats sophistication. A simple sizing rule applied on every trade outperforms an elaborate one applied when you remember.
Key takeaways
Confidence scores are only useful if they change your actions, and position sizing is the mechanism. Verify calibration first, because every formula downstream assumes it. Scale a base size by confidence, with a threshold below which you do not trade. Fold in risk/reward through expected value. Use fractional Kelly as a ceiling, never full Kelly as a target. Respect portfolio limits, because correlation makes separate positions less separate than they look.
Every Pearlixa signal ships with the inputs this framework needs: a calibrated confidence score, entry, target, and stop on every call. Early access is open at pearlixa.com/early-access.
*Cryptocurrency trading involves substantial risk of loss. This content is educational and is not investment advice. The sizing schemes described are illustrations of published methods, not recommendations for your capital.*