Page 1 of 2

Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 4:42 pm
by Shadow Trader
Twenty trades feels like a meaningful sample, especially after a bad week. In practice it says very little about whether a method works.

Take a strategy that wins 55 percent of the time with equal-sized wins and losses. Over twenty trades, the number of winners will often land anywhere from around seven to sixteen just by chance. That's the same strategy producing a record that looks like a disaster at one end and like a breakthrough at the other. Losing streaks of five or six in a row are also entirely normal for a method like that over a few hundred trades.

This matters because the most common reaction to a bad short run is to change something: tweak the entry, add a filter, switch pairs. If the run was just noise, the change wasn't needed, and it can easily make the strategy worse while the next good run makes it look like the change worked.

A more patient approach is to decide in advance how many trades a strategy gets before it is judged, often a hundred or more, and to set a separate drawdown limit at which trading pauses for review. That way, a normal bad stretch doesn't trigger a redesign, but a genuinely broken method still gets caught. How many trades do people here want to see before trusting or discarding a system?

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:23 pm
by FTtrader
Shadow Trader wrote: Fri Nov 13, 2026 4:42 pm Twenty trades feels like a meaningful sample, especially after a bad week. In practice it says very little about whether a method works.

Take a strategy that wins 55 percent of the time with equal-sized wins and losses. Over twenty trades, the number of winners will often land anywhere from around seven to sixteen just by chance. That's the same strategy producing a record that looks like a disaster at one end and like a breakthrough at the other. Losing streaks of five or six in a row are also entirely normal for a method like that over a few hundred trades.

This matters because the most common reaction to a bad short run is to change something: tweak the entry, add a filter, switch pairs. If the run was just noise, the change wasn't needed, and it can easily make the strategy worse while the next good run makes it look like the change worked.

A more patient approach is to decide in advance how many trades a strategy gets before it is judged, often a hundred or more, and to set a separate drawdown limit at which trading pauses for review. That way, a normal bad stretch doesn't trigger a redesign, but a genuinely broken method still gets caught. How many trades do people here want to see before trusting or discarding a system?
Hi Shadow trader,

This is an incredibly accurate and timely take. You’ve perfectly diagnosed one of the most common psychological traps in trading: the illusion of validity in small sample sizes.

To build on your point, when a trader evaluates a system after 20 trades, they aren't actually evaluating the strategy's statistical edge—they are only evaluating their own recent luck and emotional tolerance for variance.

Here is an extended perspective on why your approach is correct, the dangers of mid-drawdown tweaking, and a structured benchmark for when to actually trust a system.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:23 pm
by FTtrader
The Mathematics of "Normal" Drawdowns

Your example of a 55% win rate (assuming a 1:1 risk-to-reward ratio) is spot on. Human intuition is notoriously bad at grasping probability distributions. If you run a strategy with a true 55% win rate through a 1,000-trade Monte Carlo simulation, you will almost certainly encounter a streak of 7 to 9 consecutive losses at some point.

If a trader is risking 3% of their account per trade, an entirely normal 8-trade losing streak results in a 24% drawdown. To the uninitiated, the system feels broken. To the quantitative trader, the system is simply operating within its expected parameters. If you cannot survive the mathematical reality of your own win rate, the edge is irrelevant.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:23 pm
by FTtrader
The "Curve-Fitting" Trap

As you noted, the most dangerous reaction to a bad short run is to "tweak" the system. In algorithmic and quantitative trading, this is known as strategy drift or overfitting.

When you add a moving average filter or switch timeframes specifically to avoid the losses you just took over the last 20 trades, you are no longer trading a system with a proven historical edge. Instead, you have curve-fit your strategy to perfectly navigate a 20-trade sequence that has already happened and will never happen in exactly the same way again. By over-optimizing to avoid recent noise, you often choke out the strategy's ability to capture future signal.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:24 pm
by FTtrader
How Many Trades Before Trusting a System?

Answering your final question requires looking at a system through three distinct lenses rather than just a flat number:

The Statistical Minimum (100–300 Trades): From a purely mathematical standpoint, 100 trades is generally the minimum threshold to achieve a statistically significant sample size where the Law of Large Numbers begins to smooth out the variance. For high-frequency or scalping strategies, this number should be closer to 300+ before out-of-sample trust is established.

The Regime Rule (Time > Volume): 100 trades executed over two weeks of a trending bull market tell you nothing about how the strategy performs. A robust sample size must span different market regimes—low volatility, high volatility, directional trends, and choppy consolidation. Often, proving a system requires 6 to 12 months of live forward-testing, regardless of how fast you hit 100 trades.

The "Uncle Point" (Invalidation Metric): As you astutely mentioned, you need a pre-defined drawdown limit. Trusting a system doesn't mean blindly trading it into a blown account. If historical backtesting shows your maximum expected drawdown over 1,000 trades is 15%, you might set your "Uncle Point" at 20%. If the system hits a 20% drawdown, trading halts. The system isn't necessarily discarded, but it goes back to the sandbox for structural review.

Patience and statistical discipline are what separate professional risk managers from people who are just gambling with charts. Pre-defining the evaluation criteria before the first order is even placed is the only way to keep the analytical brain in control when the inevitable losing streak hits.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:31 pm
by FTtrader
To simulate your strategy's expected drawdown, you need to extract your historical trade distribution and resample it with replacement. This breaks the chronological sequence of your past trades to reveal the mathematical probability of severe losing streaks (the "left tail" risk) over thousands of alternate realities.

Because your strategy relies on an anti-martingale scaling model where position size is dynamically adjusted based on floating profits, a simple additive Monte Carlo simulation won't be accurate. You must model compounding equity dynamically.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:32 pm
by FTtrader
Here is the exact architecture to run this locally in Python using NumPy, taking advantage of your workstation's compute capacity.

Phase 1: Data Normalization
Export your closed trade history from IC Markets (MT4/5 or cTrader). Strip out the raw currency PnL and convert every trade result into a percentage return on the account balance at the time of execution.

Winner: +0.03 (3% gain)

Loser: -0.015 (1.5% loss)

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:32 pm
by FTtrader
Phase 2: The Monte Carlo Engine (Python)

This vectorized NumPy script runs 10,000 simulations of 500-trade sequences in milliseconds. It uses cumprod (cumulative product) to accurately model how your balance scales up during winning streaks and draws down during losses.

Code: Select all

import numpy as np

# 1. Load your historical trade returns (as decimal percentages)
# Example: 55% win rate, 1.5% average win, 1.2% average loss
historical_trades = np.array([0.015, -0.012, 0.015, 0.015, -0.012, -0.012, 0.015]) 

# 2. Simulation parameters
simulations = 10000       # Number of alternate realities
sample_size = 500         # Trades per simulation

# Convert returns to equity multipliers (e.g., 0.015 -> 1.015)
multipliers = 1 + historical_trades

# 3. Resample with replacement (The Monte Carlo core)
# Creates a matrix of 10,000 rows by 500 columns
mc_results = np.random.choice(multipliers, size=(simulations, sample_size), replace=True)

# 4. Calculate equity curves using cumulative product for compounding
equity_curves = np.cumprod(mc_results, axis=1)

# 5. Calculate Peak-to-Trough Drawdowns
running_max = np.maximum.accumulate(equity_curves, axis=1)
# Drawdown formula: (Peak - Current) / Peak
drawdowns = (running_max - equity_curves) / running_max

# Find the maximum drawdown in each of the 10,000 simulations
max_drawdowns_per_sim = np.max(drawdowns, axis=1)

# 6. Extract Statistical Confidence Intervals
median_dd = np.percentile(max_drawdowns_per_sim, 50)
confidence_95_dd = np.percentile(max_drawdowns_per_sim, 95)
confidence_99_dd = np.percentile(max_drawdowns_per_sim, 99)
ruin_prob = np.mean(max_drawdowns_per_sim >= 0.25) * 100 # % chance of hitting 25% DD

print(f"Median Expected Max Drawdown: {median_dd * 100:.2f}%")
print(f"95% Worst-Case Max Drawdown: {confidence_95_dd * 100:.2f}%")
print(f"99% Black Swan Max Drawdown: {confidence_99_dd * 100:.2f}%")
print(f"Probability of hitting 25% Drawdown: {ruin_prob:.2f}%")

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:33 pm
by FTtrader
Phase 3: Interpreting the Output for System Invalidation

The output dictates your "Uncle Point." If the 95th percentile worst-case maximum drawdown over 500 trades is 22%, then a live drawdown of 15% is entirely normal noise. You do not touch the cBot logic, add filters, or pause execution.

However, if your live trading hits a 24% drawdown, the system has breached the 95% confidence interval. At this point, the statistical probability that the market regime has shifted—or edge decay has occurred—is incredibly high. Trading pauses, and the M15 market structure conditions are re-audited.

Re: Why a good strategy can look terrible over twenty trades

Posted: Fri Nov 13, 2026 6:34 pm
by FTtrader
To account for serial correlation—which is critical for an anti-martingale system that scales position sizes during winning streaks—you must abandon standard independent random sampling. If a standard Monte Carlo evenly distributes your wins, it artificially inflates the survival rate of an aggressive compounding series.

To preserve the true clustering of winning and losing streaks (e.g., the consecutive losses that happen during a choppy, low-liquidity market regime), you need to upgrade the engine to use Block Bootstrapping or a Markov Chain.

Here is how to implement the most robust method, Block Bootstrapping, directly in Python.