CCMAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.TimeSeries
CCM — Convergent Cross-Mapping for detecting causation in nonlinear dynamical systems.
Inheritance: CausalDiscoveryBase<T> → TimeSeriesCausalBase<T> → CCMAlgorithm<T>
For Beginners
CCM tests causation by checking whether one variable’s history can “predict” another variable using nearest-neighbor reconstruction in delay-coordinate space. Crucially, if X causes Y, then Y’s history cross-maps to X (not the other way), which is the opposite of Granger causality’s logic.
How It Works
CCM is based on Takens’ theorem from dynamical systems theory. If X causes Y, then the shadow manifold reconstructed from Y should contain information about X, and cross-mapping accuracy should improve (converge) with longer time series.
Algorithm:
- For each pair (i,j), construct delay embedding M_j from variable j with lag tau and dimension E
- For each point in M_j, find E+1 nearest neighbors in M_j
- Compute simplex weights from distances: w_k = exp(-d_k / d_1)
- Cross-map: predict x_i(t) as weighted combination of x_i at neighbor times
- Compute correlation rho between predicted and actual x_i
- If rho converges (improves) with increasing library size L, j cross-maps i → i causes j
Reference: Sugihara et al. (2012), “Detecting Causality in Complex Ecosystems”, Science.
Constructors
| Constructor | Summary |
|---|---|
CCMAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |