For the first five months of running the system, my position sizing was embarrassingly simple: a fixed number of shares on every trade, regardless of the stock's price, volatility, or the quality of the signal. Every confirmed signal got the same allocation. A grade A signal on a highly volatile small-cap got the same position as a grade C signal on a stable large-cap. This approach had the advantage of being easy to implement and easy to track. It had the disadvantage of being logically incoherent — treating fundamentally different risk situations as if they were identical.
Changing this was one of the highest-impact improvements I made to the system, and it required confronting some assumptions about risk that I had been avoiding.
Why Fixed Lot Sizing Is a Hidden Risk Problem
The problem with fixed lot sizing isn't immediately obvious because it looks like consistency. Taking the same number of shares on every trade feels disciplined — you're not letting emotions influence how much you commit to any single trade. But consistency in share count is not the same as consistency in risk.
Consider two stocks. Stock A is priced at 15,000 KRW and has an average true range of 300 KRW — about 2% of its price. Stock B is priced at 45,000 KRW and has an average true range of 3,500 KRW — about 7.8% of its price. If you buy 100 shares of each and place your stop-loss one ATR below entry, the potential loss on Stock A is 30,000 KRW and the potential loss on Stock B is 350,000 KRW. You've taken the same nominal position in both, but the risk is more than ten times larger in Stock B.
I realized this problem concretely after reviewing a particularly bad week in month five. Two of my three losses that week came from high-ATR stocks where I had taken the same position size I used for everything else. The losses were larger than they should have been relative to my expectations — not because the trades were worse than average, but because the risk per trade was much higher than average, and I hadn't been accounting for it.
The ATR-Based Sizing Framework
The fix was to define my position size not in shares but in risk: a fixed amount of capital at risk per trade, adjusted based on each stock's ATR and the specific stop-loss distance for that trade.
The calculation works as follows. First, I decide how much capital I'm willing to lose on a single trade if the stop-loss is hit — a number I express as a fixed dollar amount or as a percentage of total capital. Let's call this the risk budget per trade. Second, I calculate the stop-loss distance for the specific trade: the difference between the entry price and the stop-loss level, which is placed one tick below the valley low. Third, I divide the risk budget by the stop-loss distance to get the number of shares.
For example, if my risk budget per trade is 150,000 KRW and the stop-loss distance on a specific trade is 500 KRW per share, I buy 300 shares. If the stop-loss distance on a different trade is 1,500 KRW per share, I buy 100 shares. In both cases, if the stop-loss is hit, I lose approximately 150,000 KRW. The position size adapts to equalize risk rather than equalizing share count.
This sounds straightforward, but implementing it required solving two practical problems I hadn't anticipated.
The First Practical Problem — Stop Distance and ATR Are Not the Same
My initial version of ATR-based sizing used the stock's ATR directly as the stop-loss distance in the calculation. This was a mistake. The ATR measures average price movement over a period, which is useful for calibrating whether a pullback is meaningful — but the actual stop-loss in this system is placed at a specific structural level (one tick below the valley low), not at a fixed ATR distance below entry.
The actual stop distance varies significantly from trade to trade depending on how deep the pullback was and where the valley low is relative to the entry price. Using the ATR as a proxy for stop distance systematically over-sized positions in trades where the actual stop was closer than one ATR and under-sized them in trades where the stop was farther. After the first month of using this approach, the volatility in my actual risk per trade was almost as high as before — I had just replaced one arbitrary sizing rule with a slightly less arbitrary one.
The correct approach was to calculate the stop distance from the actual stop-loss level for each specific trade, not from the ATR. The ATR is still used — it determines whether the pullback was deep enough to qualify as meaningful — but the sizing calculation uses the trade-specific stop distance, not the ATR itself. Once I made this correction, the actual risk per trade became much more consistent.
The Second Practical Problem — Signal Grade Should Affect Position Size
After about two months of using the volatility-adjusted framework, I noticed that I was sizing grade A and grade C signals identically. Both got the risk budget divided by the stop distance. But grade A signals had historically produced continuation moves at a meaningfully higher rate than grade C signals. Treating them identically meant I was not capitalizing on the information the grade system was providing.
I added a grade multiplier to the position sizing formula. Grade A signals receive 100% of the base risk budget. Grade B signals receive 75%. Grade C signals receive 50%. This keeps the maximum risk consistent across grades while allowing the position size to reflect signal quality.
The practical effect was that grade A trades, which had the highest continuation rate, also had the largest positions when they worked out — amplifying the good outcomes. Grade C trades, which had the lowest continuation rate, had the smallest positions — limiting the damage when they didn't work.
Over the three months since implementing the grade multiplier, the overall risk-adjusted performance improved noticeably. The improvement wasn't dramatic — the change in position sizing doesn't alter which signals work and which don't — but it shifted capital allocation toward the situations where the edge was stronger and away from the situations where the edge was weaker.
What the Data Showed After Six Months of Volatility-Adjusted Sizing
After six months of using the volatility-adjusted, grade-weighted approach, I compared the performance metrics to the five months of fixed lot sizing that preceded it.
The average loss per losing trade declined by about 22% compared to the fixed lot period. This makes sense mechanically: positions in high-volatility stocks, which had historically produced the largest losses, were now smaller.
The average gain per winning trade changed less — up about 8%. This was expected. The grade multiplier increased the size of A-grade positions, which tended to have larger average gains, but the effect was smaller than the loss reduction because A-grade signals were not uniformly the largest positions under the old system.
The maximum single-trade loss, which had been a persistent concern under fixed lot sizing — some trades in high-ATR stocks had produced losses several times larger than the average — declined significantly. Over the six months of volatility-adjusted sizing, the largest single loss was approximately 2.3 times the average loss, compared to 4.8 times under fixed lots. The tail risk contracted meaningfully.
The most important change, though, was psychological rather than financial. Under fixed lot sizing, I was never quite sure whether a loss was "a normal loss" or "a bad loss" in relative terms, because every trade had a different underlying risk level. Under volatility-adjusted sizing, I know before entering every trade approximately how much I'll lose if the stop is hit. That number is roughly the same every time. This consistency made it easier to evaluate performance objectively — a losing month under volatility-adjusted sizing tells me something meaningful about signal quality, whereas a losing month under fixed lots could reflect either poor signals or simply having taken larger positions in more volatile stocks.
Today's Investing Insight — The Kelly Criterion and Why Most Traders Should Use a Fraction of It
The Kelly Criterion, described briefly in an earlier post, calculates the theoretically optimal fraction of capital to risk on each trade given a known win rate and payoff ratio. In practice, very few traders use the full Kelly fraction for two reasons. First, the inputs — win rate and average payoff — are estimates that carry significant uncertainty, and errors in the inputs translate into errors in the sizing output. Second, the full Kelly fraction produces drawdowns that are psychologically very difficult to sustain even when the strategy is working correctly. A commonly used alternative is half-Kelly or quarter-Kelly: using a fixed fraction of the optimal Kelly size, which sacrifices some theoretical performance in exchange for meaningfully lower variance and drawdown. The volatility-adjusted approach described in this post doesn't use Kelly explicitly, but it addresses the same underlying problem — how much to risk on each trade — from a different angle: instead of computing an optimal fraction from win rate and payoff estimates, it sets a fixed dollar risk per trade and adjusts position size to match. Both approaches are solving for consistency in risk exposure; they differ in how they define what consistent means.
---
This post documents a personal journey of building and running an algorithmic trading system and reflects personal experience and perspective. Position sizing approaches described here are based on personal experience and are not a recommendation for any specific method. All investment decisions and their outcomes are the sole responsibility of the investor.
Comments
Post a Comment