Building a Short Tech Hedge from Scratch

Most of my models are long. They're looking for stocks about to move up.

But markets don't only go up. And if you're running a portfolio of long positions in tech, a 15–20% drawdown in the sector can erase months of gains in a week. I wanted a hedge on the short side — something that would perform when tech sold off.

Here's how I built it, tested it, waited for the right entry, and deployed it.

Step 1: Define What You're Trying to Catch

The setup for a short model is the mirror image of a long model.

Instead of looking for stocks about to move up +3%, you're looking for stocks about to move down. My spec:

• Universe: Technology stocks
• Target: -5% or more in 10 days
• Features: 30 — technicals, fundamentals, momentum, valuation signals
• Max positions: 20 at a time
• Confidence threshold: 55%
• Stop-loss: 5%

The model's job: scan every tech stock each morning and identify the ones most likely to drop within the next two weeks.

Step 2: Choose the Right Training Period

This is where short models get interesting.

You can't just train on any random stretch of market history. You need to train on a period where the thing you're trying to catch actually happened — significant tech drawdowns, overvaluation extremes, conditions where short setups were abundant.

I chose the 1997–2002 dot-com collapse as my primary training period. Five years of tech bubble inflation and the full crash. If any period taught a model what an overvalued tech sector looks like before it falls, it was that one.

Then — critically — I validated it on completely different periods the model never saw: 2021–2022 and 2018–2019.

If the patterns it learned from 1997–2002 still worked in 2021, when valuations were extreme again, that's evidence of a real edge. Not a curve-fit.

Step 3: Build the Mirror Model

One model wasn't enough. I wanted two, trained on different eras, looking for the same outcome. The logic: if both models agree on a stock, that's a higher-conviction short than if only one sees it.

The mirror model: trained on the 2021–2023 drawdown, backtested on the 2000–2002 dot-com collapse.

Results on the 2000–2002 backtest: +221.87% return, Sharpe 1.42, 64.7% win rate, max drawdown -19.90%. Tech dropped 57% over that same period.

One honest caveat: the 2021-trained model didn't translate as cleanly to the 2018 drawdown. That's expected — 2018 wasn't a valuation-extreme environment. The model is looking for the specific kind of overvaluation that existed in 2021. When conditions are different, it finds fewer targets. In 2018 it averaged less than one pick per day. That's not a failure — that's the model being selective. It was trained on frothy tech, and 2018 wasn't frothy enough.

Step 4: Build the Overlap Filter

With two models running, I created a third view: the overlap.

Overlap = stocks that appear in both models above 55% confidence on the same day.

Each model individually generates 17–18 picks per day. The overlap is usually 3–5 stocks. Those are the highest-conviction shorts — the ones both a dot-com-era model and a 2021-era model agree are set up to fall.

When IYW hit the right level, I was going to focus most of my capital on the overlap list.

Step 5: Wait for the Right Entry

The models were ready in mid-May. IYW (the iShares U.S. Technology ETF) was trading around $240–250.

I didn't deploy yet.

If IYW turned around right there, a 10–15% drawdown would put it roughly flat on the year. That's not enough edge. I wanted the setup more in my favor before putting real capital at risk. I set a target: IYW at $260.

For three weeks I tracked the models every day — watching their performance, watching the overlap picks, watching IYW. The models were right about a lot of individual stocks during that period. 63.6% average confidence, +4.54% average return, 100% win rate over 6 live days of tracking. The model was working. I was just waiting for the macro setup.

IYW hit $260 in early June.

Step 6: Deploy

June 3rd. The model had 38 picks. I tried to short as many as I could get borrow on — not all stocks are available for short selling at a given moment. Over the first two days I got to about 18–19% of the portfolio short. My target was 15–30%, so right in the range I wanted.

Results:

• June 3: 70.4% avg confidence, +8.55% avg return, 75% win rate
• June 4: 65.7% avg confidence, +12.55% avg return, 100% win rate

IYW dropped from $260 to $240 by Friday. A $20 drop in four days. The model called the direction.

The Honest Take

I set my price targets too low on a lot of positions.

Some exited at the stop loss (-5%) during intraday volatility, then kept going lower after I was out. The model was right. My profit target setting was too tight. I left money on the table.

The overall trade still made money. But if I had given positions more room on the target side, it would have been a much bigger week.

That's just how this works. The model handles direction. You handle execution. Some execution decisions you'd change in hindsight. I'll recalibrate the PT next time and keep the model settings the same — the model did its job.

What the 30-Day Chart Shows

Looking at 20 days of combined model data:

• 68.1% average confidence
• +4.63% average return per pick
• 53% win rate
• Model equity curve: up. QQQ over the same period: roughly flat.

Short models don't need high win rates to be profitable — each winning short more than offsets the stopped-out positions when the average win is larger than the average loss. A 53% win rate with good average returns is a working edge.

The Takeaway

Building a short model on Quant-Builder took a couple of hours across a few sessions. Training the mirror took another hour. The overlap filter is automatic — it's built into how the platform combines models.

The hard part wasn't building the model. It was having the patience to wait for the right macro setup before deploying. That part is always yours.

The model finds the setups. You decide when to act on them.

Long models running every morning. Short models ready when tech gets expensive again. Cash cycling naturally between them. That's what a complete systematic trading setup actually looks like.