Direct answer: what settings change
Fractal Dimension Index (FDI) settings change how strongly the indicator reacts to recent price movement and how much randomness it smooths away. In practical terms, settings usually influence the data segment used (for example, a lookback window), how observations are sampled, and whether the output is smoothed. Those choices change the balance between responsiveness and noise. The key trade-off is that settings that make FDI “more sensitive” also tend to produce more frequent short-term fluctuations, while settings that make it “more stable” often lag behind sudden changes.
Because an indicator is a calculation on chosen inputs, the most important verification step is consistency: use the same method and settings when comparing two periods, and be cautious when extrapolating any historical relationship. There is no universal setting that works the same across all market conditions.
Mechanism and definition: what FDI is estimating
A fractal-dimension-style indicator is intended to estimate how “complex” or “rough” a time series appears over a range of scales. The FDI idea is based on the fact that a path can look different depending on the resolution: at fine scales it may be jagged, while at coarser scales it can appear smoother. The indicator converts that intuition into a numerical value by applying a calculation to a series of observations.
Settings affect the estimation mainly by changing:
- Which observations enter the calculation: a longer lookback window includes older information and typically makes the estimate less sensitive to very recent noise.
- How that information is organized: some implementations use rescaling or multi-scale steps; settings can alter the set of scales being compared.
- How the output is processed: smoothing or averaging can reduce high-frequency variability, shifting the index toward a slower-moving curve.
These changes do not alter the underlying “definition” of complexity estimation, but they change what portion of the data and what time resolution the estimate reflects.
Evidence via example: sensitivity vs noise
Consider two hypothetical implementations that differ only in the lookback window length:
- Short window assumption: the indicator computes the estimate using mostly recent observations. If the price oscillates within a narrow range, the estimate may still swing because the algorithm is working with limited context.
- Long window assumption: the indicator computes the estimate using more history. The same oscillation may produce a smaller swing because older observations anchor the estimate.
If you also add smoothing, you can further reduce sharp swings. However, that smoothing can hide turning points because it blends past and present values. So, settings can change the shape of the index (how quickly it moves and how often it fluctuates), not a promised outcome.
A practical self-check is to run the same indicator on a time series with clearly marked regime changes (for example, a period of low volatility followed by higher volatility) while keeping all other inputs identical. You should expect faster-reacting settings to align more quickly with the change, but also to show more transient variability.
Limitations and risks: when settings can mislead
At least three material limitations should shape how you interpret “setting changes”:
- Noise amplification: more sensitive settings can make the index appear to react strongly to ordinary randomness. This can lead to misinterpretation if you treat fluctuations as meaningful by default.
- Lag and smoothing bias: more stable settings may not reflect a change until after it has progressed, because the calculation averages across time.
- Failure to transfer across conditions: historical behavior at one volatility regime or data quality level may not match another. Relationships are not guarantees.
Other common failure modes include:
- Inconsistent data choices (different sampling frequency, missing observations, or data cleaning differences) that change the computed values.
- Overfitting settings to past periods, where you accidentally find settings that match past noise.
- Execution and cost effects: even if an indicator behaves as expected on the input series, real-world outcomes can differ due to spreads, slippage, and latency. Since those are not part of the indicator computation, you cannot infer them directly from the index.
Verification and next question
To independently verify how settings change FDI, compare outputs under controlled changes: