What “responsible” CMO backtesting means
Backtesting a CMO (Chande Momentum Oscillator) means computing its historical values from price data and evaluating how some decision rule would have performed if it had been applied in real time. “Responsibly” focuses on whether the test is reproducible and whether it avoids common errors that make past results look better than they could have been.
A responsible approach is mainly about four things: (1) defining the indicator and data precisely, (2) including costs and realistic timing, (3) controlling bias and parameter overfitting, and (4) validating out of sample so you can estimate how fragile the findings are.
Mechanics: define the indicator, data, and assumptions
Before you test anything, you need a clear, verifiable definition of CMO and the inputs it uses. In practice, this means writing down:
- The exact data series (e.g., close-to-close prices) and the sampling frequency used to compute returns or momentum.
- The lookback window (the parameter) used in the CMO calculation.
- The handling of missing data, session breaks, and corporate actions (where applicable).
- The date alignment rule: whether the signal you test at time t uses only information available up to t.
Then specify any decision rule you are evaluating. Even if you are not claiming that CMO “predicts,” you must still describe what you would do when CMO crosses a level, rises by a threshold, or when a chosen condition holds. For every example, state your assumptions explicitly, such as position start time (at the next bar open vs. at bar close), whether you allow intrabar execution, and how you treat weekends or illiquid periods.
Costs and execution timing: the part many tests ignore
Costs are not an afterthought in backtesting. Even a conceptually sound indicator can appear strong if you ignore realistic friction. In a typical market simulation, you should account for:
- Spread and commissions (or an equivalent transaction cost model) and whether the cost is applied on entry only or on both entry and exit.
- Slippage: the difference between the modeled execution price and the intended price.
- Timing: the signal timestamp must not “cheat.” If your signal is computed from bar t, a realistic test uses an execution price that would have been available after bar t completes.
Assumption clarity matters: for instance, if you assume execution at the next bar open, you should apply that rule consistently and confirm that your indicator computation does not use future prices.
Bias controls: prevent look-ahead and overfitting
Many backtests fail because they accidentally use information they should not have had.
Key bias checks include:
- Look-ahead bias: ensure the indicator value used for a decision at time t is computed from data available before t.
- Selection bias: don’t tune indicator parameters (like the CMO window) by repeatedly searching for the best historical outcome on the same dataset.
- Data snooping: if you try many variants (thresholds, parameter values, filters), your results can reflect the search process rather than a robust property.
A practical control is to separate data into at least two sets: one for choosing parameters and another for evaluating performance. If you further tune multiple times, you can treat that tuning as part of the model-building process and keep the final evaluation untouched.
Evidence and out-of-sample validation
To reduce the risk of mistaking noise for structure, validate out of sample.
Common patterns include:
- Train/validation/test splits: choose parameters on a training segment, validate on a different segment, and only then do a single evaluation on the held-out test segment.
- Walk-forward (rolling) evaluation: repeatedly retrain or re-select parameters on a moving window, then test on the next period.
When reporting results, focus on metrics that make fragility visible. Use multiple measures (e.g., average return and drawdown-like risk proxies) and check whether performance depends heavily on a few unusual periods. If the test works mainly when market conditions resemble the training regime, it is likely to fail when regimes shift.
Limitations and failure modes you should expect
Even with careful methodology, CMO backtests can produce misleading conclusions. Material limitation examples include:
- Non-stationarity: markets change; relationships that held historically may not persist. - Regime shifts: volatility, trends, and liquidity conditions can alter how momentum oscillators behave.