backtesting

I tested that theory on a Bollinger Band mean-reversion system for $BTC and $ETH . The data says otherwise. 🔬
📊 Test parameters
• Assets: BTCUSDT and ETHUSDT (tested separately)
• Timeframe: 1H
• Period: Jan 2024 to Present
• Friction: 0.05% commission per trade, no slippage modelled
• Sample: 153 trades on BTCUSDT , 87 trades on $ETH (selected sets)
• Process: Optimized 192 parameter combinations on 2024 data only, then ran one continuous walk-forward test over the full history.
📐 Strategy rules
• Long: Previous 1H candle closed below the lower Bollinger Band, current candle closes back above it, and price is above a long-term EMA. (Shorts are the mirror image).
• The trend filter ensures a strong breakout is never faded head-on.
• Stop: Fixed ATR-based hard stop. No trailing.
• Target: Fixed ATR-based full exit inside the opposite band.
• Risk: Position size is a fixed fraction of equity divided by stop distance.
⚠️ Reality check
The flattering part: The selected sets were net-positive in every single year. returned +15.34% (max drawdown -7.56%). ETHUSDT returned +17.37% (max drawdown -5.18%).
The part that hurts: Those sets were chosen after seeing the full history. That isn't clean out-of-sample data. When I judged every set that was profitable in 2024 only on the later blind years:
Only 10.59% of BTCUSDT sets and 12.77% of ETHUSDT sets stayed positive in every out-of-sample year.
The median out-of-sample return fell to -2.89% for BTCUSDT and -2.24% for ETHUSDT.
The worst-case scenario was -17.68% for BTCUSDT and -20.69% for ETHUSDT.
A better training year did not predict a better future (Rank correlation was negative for both assets).
Exactly 0 sets met my quality targets for out-of-sample profit factor and drawdowns.
🛠 How I handle it
Why it fails: Mean reversion earns small, frequent wins, but gets crushed when price trends heavily through the bands. The 2024 optimization mostly just memorized that specific year's market chop. It didn't carry forward.
What I do about it: I keep the trend-bias filter, add range-regime and volatility filters so the system stays entirely out of trending conditions, and fix risk per trade with ATR-based stops. I judge sets on their worst year, not their best total return.
This is research, not a promise of returns. If you want to see how a systematic approach actually behaves live—drawdowns included—follow my Binance lead-trader profile. 📉
🔗 Lead Trader Profile
(Disclaimer: Backtested and past results are hypothetical and do not guarantee future performance. Not financial advice.)
#cryptotrading #Backtesting #QuantTrading #bitcoin

I ran an EMA trend + RSI pullback system on $BTC and $ETH and got exactly that. Then I checked how much of it was just selection bias. Here's the data. 👇
📊 Test parameters
• Assets: BTCUSDT and ETHUSDT (tested separately, $1M starting cash each)
• Timeframe: 1H
• Period: 2024 to 2026, one continuous backtest
• Friction: 0.05% commission per trade
• Optimized: 120 parameter combinations tested on 2024 only, winners run over the full history.
• Acceptance rule: Net-positive in every calendar year.
📐 Strategy rules
• Long/Short: Price is on the correct side of a long-term EMA, and RSI crosses back through a pullback/rally level.
• Filter: Volume must be at or above its recent average (entries only happen with participation).
• Risk & Exits: ATR-based hard stop. Half position closed at ATR target, stop moves to break-even, the rest trails. Position size strictly capped based on equity.
⚠️ Reality check
The flattering part: The accepted sets were net-positive in every single year.
• $BTC : +31.78% return, Sharpe 1.48, max drawdown -7.37% (182 trades).
• $ETH : +36.26% return, Sharpe 1.12, max drawdown -9.21% (262 trades).
The part that hurts: That "every year positive" filter was applied to the full history, meaning it relied on hindsight. For a cleaner test, I took every set that was profitable in 2024 and judged it only on the blind 2025-2026
$BTC : 110 sets were profitable in 2024.
• Only 43.6% stayed positive in later years.
• Only 7.2% kept a profit factor of at least 1.3.
• The worst out-of-sample total return was -17.13%.
$ETH : 47 sets were profitable in 2024.
• Only 38.3% stayed positive in later years.
• Exactly 0.00% held a profit factor of at least 1.3.
Rank correlation between 2024 returns and later returns: -0.06 on $BTC , -0.12 on$ETH . A better 2024 told me absolutely nothing about what came next.
🛠 How I handle it
Why does it break? A pullback-recovery entry assumes the trend will resume. When the market chops or the regime shifts, the signal fires and the move simply fades.
Fitting one year gave no edge in picking the next year's winner. So, I don't trust the top optimized row. I rank candidates by their worst year instead of total return, and I strictly enforce the trend and volume filters so a bad stretch stays contained. The chosen sets still carry in-sample bias, and I'd rather admit that than hide it. Backtests are not guarantees, and past results don't promise future ones.
If you value rigor over hype, follow my Binance lead-trader profile and judge the live, risk-managed results yourself. 📊
🔗 Lead Trader Profile
#bitcoin #Ethereum #strategy #Backtesting

A strategy needs clear rules so that its results can be analyzed consistently.
When parameters, criteria, or conditions are changed during testing, the result no longer represents exactly the same method. This makes it difficult to know whether the performance came from the original strategy or from the changes made along the way.
This problem appears frequently when someone adjusts a strategy after observing certain results and then tests again using the new rules. The result may seem better, but the comparison is no longer happening under the same conditions.

The backtest promised my hourly volume alerts would hit +10% within 12 hours—28.72% of the time.
The machine has been running for real. 170 alerts closed the window. Score:
**23.46%** on 162 clean orders.
Lower than the backtest by 5.26 points. Just as in every backtest: the real trade performs worse than the “paper” one.
But still **3.56 times** better than a random hour (6.59%). The edge is real, just smaller than the advertisement’s claim.
Three things that come with that number.
**One.** 8 warnings are still open and **are not counted**. A position not finished within the time window isn’t a win.

In my confessions as a rookie trader, today I want to tell you why trading without a system is like driving blind...
I must confess that when I started in this game, one of the hardest things for me was remembering that an RSI of 70 meant overbought and 30 meant oversold. Something so 'basic' created a terrifying void in my mind. And no, it wasn't due to a lack of intelligence; it's just that trading with dyslexia and dyscalculia forces you to live in a universe where numbers and directions sometimes turn into a dead-end tunnel.