The standard disclaimer is not boilerplate
Every financial product says it: "Past performance is not indicative of future results." Most readers skip over it. This is a mistake — the statement captures one of the most reliable findings in quantitative finance, and its implications are severe for anyone evaluating a trading strategy.
Overfitting: the silent killer of backtests
A backtest is a simulation of how a strategy would have performed had it been applied to historical data. The problem is that the strategy designer — consciously or not — has seen that historical data while building the strategy. They have, at minimum, a mental model of what the chart looked like.
This creates overfitting: the strategy is tuned, implicitly or explicitly, to the specific characteristics of the historical period. When applied to new data (which has different characteristics by definition), it underperforms.
The more complex the strategy and the more parameters it has, the worse this problem gets. A strategy with 20 tunable parameters and 24 months of training data is almost certainly overfit.
Coinblockers' own version of this problem
The Coinblockers neural network validator is trained on the site's own historical signal outcomes. This creates exactly the overfitting risk described above. The training set covers a specific market regime (crypto 2023–2025, post-FTX recovery into the 2024 halving bull run). If the next period is a sideways grind or a deep bear market with different correlation structure, the model's learned patterns may not transfer.
We publish this caveat explicitly because we know most users won't read this article. The model is updated when enough new labeled outcomes accumulate, and the validation statistics are shown on the model health admin page.
Regime change is the main enemy
Financial time series are not stationary. A strategy that works brilliantly during high-volatility trending markets will typically fail during low-volatility ranging markets, because the patterns it exploits don't exist in the same form.
Crypto markets have experienced several distinct regimes in the past five years: the 2020-2021 retail-driven mania, the 2022 institutional deleveraging, the 2023-2024 post-crash accumulation, and the 2024-2025 ETF-inflow bull market. A strategy trained on any two of these will be surprised by the third.
What "educational use only" actually means
When Coinblockers says its signals are for educational purposes only, this is not just legal cover. It reflects a genuine epistemic claim: we do not know whether these signals will generate positive returns in the next 12 months, and we have no way to test that claim in advance.
What we can offer is transparency: the full methodology, the historical virtual track record with all losing trades included, and the real-time paper trading engine that generates new data continuously. You can evaluate the methodology on its own terms and decide whether it's plausible.
The right response to a backtest showing a 40% return is: "What assumptions went into this, and do I believe those assumptions will hold going forward?" Not: "This is going to return 40%."
Practical takeaways
- Demand live paper-trading data, not just backtests. Backtests can be constructed to look good; paper-trading in real time cannot.
- Longer live track records are worth exponentially more than longer backtests with the same statistical performance.
- Underperformance relative to a backtest is the norm, not the exception, especially in the first 12-24 months of live operation.
- Complexity is not a feature — simpler strategies overfit less and generalize better across regimes.