feat: state-space seasonal model matching R bsts AddSeasonal()

.cargo/config.toml

@@ -3,6 +3,3 @@ rustflags = ["-C", "target-cpu=native"]
 
 [target.aarch64-apple-darwin]
 rustflags = ["-C", "target-cpu=native"]
-
-[target.x86_64-unknown-linux-gnu]
-rustflags = ["-C", "target-cpu=native"]

.gitignore

@@ -13,3 +13,4 @@ z-ai/
 .claude/
 *.DS_Store
 site/
+.cargo/

README.md

@@ -103,7 +103,7 @@ Tests run on every commit with seed-fixed MCMC for deterministic reproduction.
 |---|---|---|
 | `point_effect_mean` | ±3% relative | Passing on core scenarios |
 | `cumulative_effect_total` | ±3% relative | Passing on core scenarios |
-| `ci_lower` / `ci_upper` | Tight parity | `±1.5%` no-covariates, `±1%` covariates, explicit Phase 2 acceptance `±3%`, seasonal fixture `±5%` |
+| `ci_lower` / `ci_upper` | Tight parity | `±1%` no-covariates, `±1%` covariates, explicit Phase 2 acceptance `±3%`, seasonal `±1%` |
 | `p_value` | Significance match | Classification at alpha=0.05 |
 
 ### What is matching R and what is not
@@ -118,10 +118,13 @@ Tests run on every commit with seed-fixed MCMC for deterministic reproduction.
 | Spike-and-slab variable selection | Matching | Coordinate-wise sampling with StudentSpikeSlabPrior defaults (`expected.r2=0.8`, `prior.df=50`, `prior.information.weight=0.01`, `diagonal.shrinkage=0.5`) |
 | expected.model.size | Matching | Unified default `2` in `CausalImpact` and `ModelOptions` |
 | expected.r2 = 0.8, prior.df = 50 | Matching | Same documented residual variance prior defaults as BoomSpikeSlab / bsts |
-| Seasonal component (`nseasons`, `season_duration`) | Supported | R-compatible API with seasonal fixture coverage |
+| Seasonal component (`nseasons`, `season_duration`) | Matching | State-space model matching R bsts `AddSeasonal()` (±1% CI parity) |
 | Dynamic regression | Supported | Time-varying coefficients via random-walk FFBS; `dynamic_regression=True` |
 | Local linear trend | Supported | Opt in with `state_model="local_linear_trend"` |
 
+Matching = CI-enforced numerical equivalence with R bsts (±3% or tighter).
+Supported = Feature implemented, no R parity fixture yet.
+
 Covariate CI bounds are enforced twice: the legacy parity fixture remains tighter than
 Phase 2 requirements, and a separate Phase 2 acceptance test keeps the threshold at `±3%`.
 

docs/api.md

@@ -56,8 +56,8 @@ ci = CausalImpact(data, pre_period, post_period, model_args=opts)
 | `expected_model_size` | `int` | 2 | Expected number of active covariates for spike-and-slab prior |
 | `dynamic_regression` | `bool` | `False` | Enable time-varying regression coefficients |
 | `state_model` | `str` | `"local_level"` | `"local_level"` or `"local_linear_trend"` |
-| `nseasons` | `int \| None` | `None` | Seasonal cycle count |
-| `season_duration` | `int \| None` | `None` | Duration of each seasonal block; defaults to 1 when `nseasons` is set |
+| `nseasons` | `int \| None` | `None` | Seasonal cycle count. `nseasons=1` is equivalent to no seasonal component. |
+| `season_duration` | `int \| None` | `None` | Duration of each seasonal block; defaults to 1 when `nseasons` is set. Requires `nseasons` to be set. |
 
 ## `CausalImpactResults`
 

docs/compatibility-matrix.md

@@ -8,7 +8,7 @@ Comparison of features between R CausalImpact (bsts 1.4.1) and this Python imple
 |---|---|---|---|
 | Local level | Yes | Yes | Identical algorithm |
 | Local linear trend | Yes | Yes | `state_model="local_linear_trend"` |
-| Seasonality | Yes | Yes | R-compatible API with seasonal fixture coverage |
+| Seasonality | Yes | Yes | State-space model matching R bsts `AddSeasonal()` |
 | Dynamic regression | Yes | Yes | `dynamic_regression=True` |
 | Regression (static) | Yes | Yes | Identical algorithm |
 
@@ -18,7 +18,7 @@ Comparison of features between R CausalImpact (bsts 1.4.1) and this Python imple
 |---|---|---|---|
 | niter | Yes | Yes | Same default (1000) |
 | nseasons | Yes | Yes | `ModelOptions.nseasons` or `model_args["nseasons"]` |
-| season.duration | Yes | Yes | `ModelOptions.season_duration` or `model_args["season.duration"]` |
+| season.duration | Yes | Yes | `ModelOptions.season_duration` or `model_args["season_duration"]` (R compat: `"season.duration"`) |
 | prior.level.sd | Yes | Yes | Same default (0.01) |
 | standardize.data | Yes | Yes | Same default (True) |
 | expected.model.size | Yes | Yes | Unified default `2` |
@@ -71,7 +71,7 @@ Comparison of features between R CausalImpact (bsts 1.4.1) and this Python imple
 |---|---|---|
 | point_effect_mean | ±3% relative | Passing |
 | cumulative_effect_total | ±3% relative | Passing |
-| ci_lower / ci_upper | Tight parity (`±1.5%` no-cov, `±1%` covariates, `±5%` seasonal) | Passing |
+| ci_lower / ci_upper | Tight parity (`±1%` no-cov, `±1%` covariates, `±1%` seasonal) | Passing |
 | p_value significance | Match at alpha=0.05 | Passing |
 
 Tests run against R CausalImpact 1.4.1 fixtures on every PR.

docs/examples.md

@@ -183,6 +183,38 @@ Posterior prob. of a causal effect: 99.95%
 True effect = 8.0, estimated = 7.57. The 95% CI [5.99, 9.18] contains the true
 value.
 
+### How it works
+
+The seasonal component uses a state-space model matching R bsts `AddSeasonal()`.
+The state vector is `[μ_t, s_1(t), ..., s_{S-1}(t)]` where `S = nseasons`.
+Seasonal states evolve via a sum-to-zero transition: the next season equals
+the negative sum of the previous `S-1` seasons plus noise.
+
+### Using `season_duration`
+
+Set `season_duration > 1` when each season spans multiple time steps.
+The seasonal transition only fires at season boundaries
+(`t % season_duration == 0`); in between, the seasonal state is frozen.
+
+```python
+# Using the same data from the seasonal example above
+
+# Daily data with 7-day weekly pattern (default: season_duration=1)
+ci = CausalImpact(
+    data, pre_period, post_period,
+    model_args={"nseasons": 7, "season_duration": 1, "seed": 42},
+)
+
+# Monthly data with quarterly pattern (each quarter = 3 months)
+ci = CausalImpact(
+    data, pre_period, post_period,
+    model_args={"nseasons": 4, "season_duration": 3, "seed": 42},
+)
+```
+
+When `season_duration` is omitted, it defaults to 1 (every time step is a new
+season). `nseasons=1` is equivalent to no seasonal component.
+
 ---
 
 ## 4. Dynamic Regression

docs/migration-from-r.md

@@ -81,7 +81,9 @@ Key differences:
 
 ## Numerical Equivalence
 
-This library verifies ±3% agreement with R CausalImpact on point estimates and cumulative effects across multiple test scenarios, including a seasonal fixture. Tests run on every PR.
+This library verifies ±3% agreement with R CausalImpact on point estimates and cumulative effects across multiple test scenarios. Tests run on every PR.
+
+The seasonal model uses the same state-space algorithm as R bsts `AddSeasonal()`, achieving ±1% CI parity on the seasonal fixture.
 
 Differences arise from independent RNG implementations (R's `set.seed` vs Rust's `ChaCha8Rng`), not from algorithmic differences.