The Backtest Curve Looked Great. I Cut the Position Anyway.

Last Saturday around 11pm, I was sitting at my desk tuning parameters for an automated prediction system I've been writing for over two years now. Every day after market close, it runs through historical data and decides how much to bet the next day. That night I was staring at a backtest curve that kept climbing, my cursor hovering over the "position multiplier" field, debating whether to double it across the board. The curve looked genuinely tempting: daily compound returns jumping from just over 3% to around 11%.

The problem this system solves is simple: I can't watch the market and calculate odds every day myself, so I wrote a program to do it for me. It finds patterns in historical data and works out how big a bet to place each time, all in service of staying calm and not letting emotion drive decisions. The "should I raise the overall position" question comes up because backtesting software is brutally honest — crank the parameter up, and it hands you a better-looking return number immediately, no hiding.

That night I nearly bumped the multiplier up right then. What stopped me was one more number I glanced at: worst-month drawdown. With the position doubled across the board, the single worst month in the historical data would have wiped out close to 97% of the account. Not a paper loss — a month that could burn the principal down to almost nothing. That 11% daily return was bought with that risk, and backtesting software will never print that sentence in bold across the screen. It just draws the pretty line a little thicker.

This wasn't the first time I'd gotten fooled by staring at an average. In the two years I've spent building this system, I've adjusted these parameters more than once, and every time the return curve climbed I wanted to push it up another notch, until one live run actually hit a real drawdown and I started forcing myself to check the worst case before the average. Same thing this time: I dialed the multiplier back and instead kept the total risk budget fixed, reallocating it based on how reliable each signal was. Strong signal, bigger bet. Weak signal, smaller bet. Total risk exposure locked.

After that change, the daily compound return came out to 3.193%, slightly higher than the 3.105% from the no-leverage version, while the worst-month drawdown dropped from 68.3% to 59.1%. Put those numbers side by side and it's almost funny: the 11% curve looked the most aggressive, but the actual money you'd walk away with was less than the version that just managed risk properly. What decides whether you should add leverage isn't how high the average return can go, it's whether you can survive the worst single month. If you can't survive it, your account hits zero, and none of those good days after that ever happen for you.

I know this logic isn't something I invented. Quant circles have called it position sizing, or some variant of the Kelly criterion, for decades. But what actually drove the lesson home wasn't a book, it was those few minutes that night when my hand almost slipped and hit the multiplier, and the drawdown number yanked it back. Everyone's heard the theory. Whether you can actually stop yourself when the temptation is right in front of you is a different matter.

I use this same check outside the system now. Whether it's deciding to go all-in on a side project, or dumping an entire budget into one marketing push, I ask the same question first: if this turns out to be the worst case, can I survive it? Only once the answer is yes do I ask how much I could make. If the answer is no, it doesn't matter how good the return looks, I stay away.

Have you actually run the numbers on the worst case for whatever you're doing right now?