Define what you are backtesting
A backtest is a simulation that applies a fixed set of rules to historical market data to produce hypothetical outcomes. For “ADX strategies,” the essential first step is to define the indicator’s role inside the rules. For example, specify exactly how the Average Directional Index (ADX) is computed conceptually in your framework, what threshold(s) or regimes (e.g., “trend strength”) the strategy uses, and how those conditions translate into position decisions.
Two common sources of confusion are: (1) mixing indicator interpretation with trading rules (an indicator by itself is not a complete strategy), and (2) leaving the rules implicit (different interpretations lead to different results). A responsible approach starts by writing down: entry logic, exit logic, position sizing method (even if it is constant), and how you handle no-signal periods.
Separate stable mechanics from variable conditions
Treat strategy mechanics as the stable part and treat market/provider conditions as variable. Costs and execution are often the biggest variable drivers. Even if you do not assume real-time data, you can still model transaction effects conceptually: include spreads (or an equivalent cost), commissions, and slippage assumptions as inputs to your simulation.
State assumptions for every calculation or example. For instance, if you assume that orders fill at the next bar’s open, that assumption affects outcomes. If you assume a fixed slippage amount, document it and test sensitivity. Cost sensitivity checks help prevent the common failure mode where a strategy looks profitable only because trading frictions were ignored.
Control bias with disciplined experiment design
Responsible backtesting also means controlling “bias,” where the backtest setup accidentally makes results look better than they would be live.
Key bias controls you can apply without relying on any specific provider:
- Walk-forward validation: fit any thresholds or parameters only on an earlier segment, then test on a later segment.
- Out-of-sample testing: keep a final period unseen during any tuning.
- Parameter stability checks: verify that results do not collapse when parameters shift slightly.
- Multiple regimes and multiple time windows: test across different volatility/trend conditions rather than a single period.
- Data leakage prevention: ensure that signals use only information available at the decision time (for example, avoid using future bars to decide the present).
These controls address a material failure mode: overfitting. Overfitting happens when rules match historical noise instead of durable structure, so the backtest may perform well in one window but deteriorate elsewhere.
Include evidence checks and compare against alternatives
Evidence in a backtest is more credible when you look beyond a single headline metric. Evaluate whether returns (or any performance measure you choose) remain qualitatively consistent under alternative, reasonable settings.
A practical method is to define a few scenarios and rerun the backtest each time:
- Use a range of assumed execution costs (e.g., conservative vs. moderate).
- Change time granularity assumptions (at minimum, verify that results are not entirely dependent on one bar interval).
- Compare against a simple baseline such as “hold flat” (no trades) or a generic trend filter (conceptually) to ensure the indicator rules add decision value.
Even then, outcomes vary with market conditions, costs, execution, and the particular instrument characteristics you model. Historical relationships do not guarantee future results.
Limitations and a “responsible” stop condition
At least one material limitation should be part of your analysis: regime change. ADX-based interpretations often rely on trend strength behavior; if markets shift toward range-bound or differently structured dynamics, the strategy logic may stop working.
Another failure mode is measurement and implementation mismatch: the way you compute or sample the indicator and apply decision timing can differ from how you would trade in practice. A responsible backtest therefore includes a clear mapping from “indicator condition becomes a decision” to the exact bar/time used for execution.
You should also define a stop condition for continued tuning. For example, if performance depends heavily on one narrow period, or if small parameter changes flip the result direction, treat that as evidence that the setup may not generalize.
Verification and next question to answer
To verify your understanding, be able to answer these independently:
- What are the exact rule definitions that turn ADX conditions into trade decisions?