rustima 0.1.0__tar.gz

rustima-0.1.0/.env.example

@@ -0,0 +1,16 @@
+# Template for rustima/.env — copy to `.env` and fill in real values.
+# `.env` is gitignored; never commit secrets.
+
+# ─── Test PyPI ─────────────────────────────────────────────────────────────
+# Get a token at: https://test.pypi.org/manage/account/token/
+#   - Scope: "Entire account" for the first upload (no project exists yet),
+#     then re-issue a project-scoped token after the project is created.
+TWINE_USERNAME=__token__
+TWINE_PASSWORD=pypi-AgEN...replace-with-real-token...
+TWINE_REPOSITORY_URL=https://test.pypi.org/legacy/
+
+# ─── PyPI (production) ─────────────────────────────────────────────────────
+# Uncomment when ready to publish to real PyPI.
+# Prefer GitHub Trusted Publishing via release.yml over local uploads.
+# PYPI_USERNAME=__token__
+# PYPI_TOKEN=pypi-AgEN...replace-with-real-token...

rustima-0.1.0/.github/workflows/ci.yml

@@ -0,0 +1,64 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  rust-test:
+    name: Rust Tests
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-action@stable
+
+      - name: Cache cargo registry
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cargo/registry
+            ~/.cargo/git
+            target
+          key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}
+
+      - name: Run Rust tests
+        run: cargo test --all-targets
+
+  python-test:
+    name: Python Tests (Tier A + Contract + Smoke)
+    runs-on: ubuntu-latest
+    needs: rust-test
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-action@stable
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+
+      - name: Create venv and install deps
+        run: |
+          uv venv .venv
+          uv pip install numpy pytest statsmodels scipy pandas
+
+      - name: Build wheel and install
+        run: |
+          uv pip install maturin
+          CARGO_TARGET_DIR=target_wheel uv run maturin build --out /tmp/wheels
+          uv pip install --force-reinstall /tmp/wheels/rustima-*.whl
+
+      - name: Run Python tests (excluding nightly)
+        run: .venv/bin/python -m pytest python_tests/ -m "not nightly" -v --tb=short

rustima-0.1.0/.github/workflows/nightly.yml

@@ -0,0 +1,83 @@
+name: Nightly
+
+on:
+  schedule:
+    - cron: "0 3 * * *"  # 03:00 UTC daily
+  workflow_dispatch:  # manual trigger
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  tier-b-tests:
+    name: Tier B Matrix Tests
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-action@stable
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+
+      - name: Create venv and install deps
+        run: |
+          uv venv .venv
+          uv pip install numpy pytest statsmodels scipy pandas
+
+      - name: Build wheel and install
+        run: |
+          uv pip install maturin
+          CARGO_TARGET_DIR=target_wheel uv run maturin build --out /tmp/wheels
+          uv pip install --force-reinstall /tmp/wheels/rustima-*.whl
+
+      - name: Run Tier B nightly tests
+        run: .venv/bin/python -m pytest python_tests/test_matrix_tier_b.py -v --tb=short
+
+  benchmark:
+    name: Performance Benchmark
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install Rust toolchain
+        uses: dtolnay/rust-action@stable
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+
+      - name: Create venv and install deps
+        run: |
+          uv venv .venv
+          uv pip install numpy pytest statsmodels scipy pandas
+
+      - name: Build wheel (release) and install
+        run: |
+          uv pip install maturin
+          CARGO_TARGET_DIR=target_wheel uv run maturin build --release --out /tmp/wheels
+          uv pip install --force-reinstall /tmp/wheels/rustima-*.whl
+
+      - name: Run Rust benchmarks
+        run: cargo bench --bench bench_fit -- --output-format bencher | tee bench_results.txt
+
+      - name: Run Python benchmark comparison
+        run: .venv/bin/python python_tests/bench_comparison.py | tee python_bench_results.txt
+
+      - name: Upload benchmark results
+        uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-results
+          path: |
+            bench_results.txt
+            python_bench_results.txt

rustima-0.1.0/.github/workflows/release.yml

@@ -0,0 +1,242 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - "v*"
+  workflow_dispatch:
+    inputs:
+      publish_to_pypi:
+        description: "Publish to PyPI (real, not test.pypi)"
+        type: boolean
+        default: false
+
+permissions:
+  contents: write
+
+env:
+  CARGO_TERM_COLOR: always
+
+jobs:
+  # ----------------------------------------------------------------------------
+  # Linux: manylinux2014 (glibc 2.17+) — required for PyPI acceptance
+  # ----------------------------------------------------------------------------
+  build-linux:
+    name: Linux ${{ matrix.target }} (py${{ matrix.python }})
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        target: [x86_64]
+        python: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+
+      - name: Build wheel (manylinux2014)
+        uses: PyO3/maturin-action@v1
+        with:
+          target: ${{ matrix.target }}
+          args: --release --out dist --interpreter ${{ matrix.python }}
+          manylinux: "2014"
+          working-directory: .
+
+      - name: Install wheel & smoke-test
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install numpy polars pytest
+          python -m pip install --force-reinstall --no-index --find-links dist/ rustima
+          python -m pytest python_tests/test_smoke.py -v --tb=short
+
+      - name: Upload wheel
+        uses: actions/upload-artifact@v4
+        with:
+          name: wheel-linux-${{ matrix.target }}-py${{ matrix.python }}
+          path: dist/*.whl
+
+  # ----------------------------------------------------------------------------
+  # macOS: arm64 (Apple Silicon) + x86_64 (Intel)
+  # ----------------------------------------------------------------------------
+  build-macos:
+    name: macOS ${{ matrix.target }} (py${{ matrix.python }})
+    runs-on: ${{ matrix.runner }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - runner: macos-latest
+            target: aarch64-apple-darwin
+          - runner: macos-13
+            target: x86_64-apple-darwin
+        python: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+
+      - name: Build wheel
+        uses: PyO3/maturin-action@v1
+        with:
+          target: ${{ matrix.target }}
+          args: --release --out dist --interpreter ${{ matrix.python }}
+          working-directory: .
+
+      - name: Install wheel & smoke-test
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install numpy polars pytest
+          python -m pip install --force-reinstall --no-index --find-links dist/ rustima
+          python -m pytest python_tests/test_smoke.py -v --tb=short
+
+      - name: Upload wheel
+        uses: actions/upload-artifact@v4
+        with:
+          name: wheel-macos-${{ matrix.target }}-py${{ matrix.python }}
+          path: dist/*.whl
+
+  # ----------------------------------------------------------------------------
+  # Windows: x86_64
+  # ----------------------------------------------------------------------------
+  build-windows:
+    name: Windows ${{ matrix.target }} (py${{ matrix.python }})
+    runs-on: windows-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        target: [x64]
+        python: ["3.10", "3.11", "3.12", "3.13"]
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python ${{ matrix.python }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python }}
+          architecture: ${{ matrix.target }}
+
+      - name: Build wheel
+        uses: PyO3/maturin-action@v1
+        with:
+          target: ${{ matrix.target }}
+          args: --release --out dist --interpreter ${{ matrix.python }}
+          working-directory: .
+
+      - name: Install wheel & smoke-test
+        shell: bash
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install numpy polars pytest
+          python -m pip install --force-reinstall --no-index --find-links dist/ rustima
+          python -m pytest python_tests/test_smoke.py -v --tb=short
+
+      - name: Upload wheel
+        uses: actions/upload-artifact@v4
+        with:
+          name: wheel-windows-${{ matrix.target }}-py${{ matrix.python }}
+          path: dist/*.whl
+
+  # ----------------------------------------------------------------------------
+  # sdist (source distribution) — required so non-wheel platforms can build
+  # ----------------------------------------------------------------------------
+  build-sdist:
+    name: Build sdist
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Build sdist
+        uses: PyO3/maturin-action@v1
+        with:
+          command: sdist
+          args: --out dist
+          working-directory: .
+
+      - name: Upload sdist
+        uses: actions/upload-artifact@v4
+        with:
+          name: sdist
+          path: dist/*.tar.gz
+
+  # ----------------------------------------------------------------------------
+  # GitHub Release with all wheels + sdist
+  # ----------------------------------------------------------------------------
+  create-release:
+    name: Create GitHub Release
+    needs: [build-linux, build-macos, build-windows, build-sdist]
+    if: startsWith(github.ref, 'refs/tags/v')
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Download all artifacts
+        uses: actions/download-artifact@v4
+        with:
+          path: dist/
+          pattern: "wheel-*"
+          merge-multiple: true
+
+      - name: Download sdist
+        uses: actions/download-artifact@v4
+        with:
+          name: sdist
+          path: dist/
+
+      - name: Create Release
+        uses: softprops/action-gh-release@v2
+        with:
+          files: |
+            dist/*.whl
+            dist/*.tar.gz
+          generate_release_notes: true
+
+  # ----------------------------------------------------------------------------
+  # PyPI publish via Trusted Publishing (OIDC, no API token required)
+  #
+  # Setup (one-time): on pypi.org → project settings → "Publishing" →
+  # "Add a new pending publisher" with:
+  #   owner = zongseung
+  #   repo  = Rust-python-arima
+  #   workflow = release.yml
+  #   environment = pypi
+  # Then create a GitHub environment named "pypi" with required reviewers
+  # if you want manual approval before publish.
+  # ----------------------------------------------------------------------------
+  publish-pypi:
+    name: Publish to PyPI
+    needs: [build-linux, build-macos, build-windows, build-sdist]
+    if: startsWith(github.ref, 'refs/tags/v') || (github.event_name == 'workflow_dispatch' && inputs.publish_to_pypi)
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/rustima
+    permissions:
+      id-token: write
+    steps:
+      - name: Download all wheel artifacts
+        uses: actions/download-artifact@v4
+        with:
+          path: dist/
+          pattern: "wheel-*"
+          merge-multiple: true
+
+      - name: Download sdist
+        uses: actions/download-artifact@v4
+        with:
+          name: sdist
+          path: dist/
+
+      - name: List artifacts
+        run: ls -lh dist/
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          packages-dir: dist/
+          skip-existing: true

rustima-0.1.0/.gitignore

@@ -0,0 +1,22 @@
+*.cache
+*.claude
+planner
+__pycache__
+.venv
+/planner
+/target_wheel
+**/target_wheel/
+Rust-python-arima/sarimax_rs/target/
+CLAUDE.md
+# 컴파일된 확장 모듈 (빌드 아티팩트) — .py 소스는 트래킹 유지
+python/rustima/*.so
+python/rustima/__pycache__/
+
+# Secrets / credentials (PyPI tokens, etc.) — NEVER commit
+.env
+.env.*
+!.env.example
+!.env.template
+
+# Build distribution
+/dist/

rustima-0.1.0/BENCHMARK_SPEC.md

@@ -0,0 +1,219 @@
+# sarimax-rs 성능 및 정확도 검증 기획서
+
+## 1. 목적
+
+sarimax-rs (Rust 엔진)가 Python statsmodels SARIMAX 대비 **모든 모델 유형에서** 속도 우위를 달성하고, 수치 정확도가 허용 범위 내에 있음을 체계적으로 검증한다.
+
+## 2. 검증 대상 모델
+
+| # | 모델 | order | seasonal_order | 데이터 길이 | 복잡도 |
+|---|------|-------|----------------|------------|--------|
+| 1 | AR(1) | (1,0,0) | (0,0,0,0) | 200 | Low |
+| 2 | AR(2) | (2,0,0) | (0,0,0,0) | 300 | Low |
+| 3 | MA(1) | (0,0,1) | (0,0,0,0) | 200 | Low |
+| 4 | ARMA(1,1) | (1,0,1) | (0,0,0,0) | 300 | Medium |
+| 5 | ARMA(2,2) | (2,0,2) | (0,0,0,0) | 400 | Medium |
+| 6 | ARIMA(1,1,1) | (1,1,1) | (0,0,0,0) | 300 | Medium |
+| 7 | ARIMA(2,1,2) | (2,1,2) | (0,0,0,0) | 400 | High |
+| 8 | SARIMA(1,0,0)(1,0,0,4) | (1,0,0) | (1,0,0,4) | 200 | Medium |
+| 9 | SARIMA(1,1,1)(1,1,1,12) | (1,1,1) | (1,1,1,12) | 300 | High |
+| 10 | SARIMA(2,1,1)(1,1,1,12) | (2,1,1) | (1,1,1,12) | 500 | Very High |
+
+## 3. 벤치마크 측정 항목
+
+### 3.1 속도 (Performance)
+
+| 측정 항목 | 방법 | 기대 결과 |
+|-----------|------|-----------|
+| Single fit 시간 | best-of-3, wall clock (ms) | Rust > 1.0x speedup (모든 모델) |
+| Kalman loglike 시간 | best-of-5, wall clock (ms) | Rust > 2.0x speedup |
+| Batch fit (50 series) | best-of-2, wall clock (ms) | Rust > 5.0x speedup (Rayon 병렬) |
+| Forecast (10-step) | best-of-5, wall clock (ms) | Rust > 2.0x speedup |
+
+### 3.2 정확도 (Accuracy)
+
+| 측정 항목 | 허용 범위 | 비교 대상 |
+|-----------|-----------|-----------|
+| Log-likelihood (oracle params) | \|err\| < 1e-6 | statsmodels loglike |
+| Fitted params | max \|err\| < 1e-2 | statsmodels fit params |
+| Fitted loglike | \|err\| < 3.0 | statsmodels fit loglike |
+| AIC / BIC | \|err\| < 6.0 | statsmodels AIC/BIC |
+| Forecast mean (10-step) | max \|err\| < 1e-4 | statsmodels forecast |
+| Forecast CI | max \|err\| < 1e-3 | statsmodels CI |
+
+### 3.3 수렴성 (Convergence)
+
+| 측정 항목 | 기대 결과 |
+|-----------|-----------|
+| 수렴률 (converged=true) | >= 90% (모든 모델) |
+| 반복 횟수 비교 | Rust <= statsmodels * 1.5 |
+
+## 4. 테스트 실행 절차
+
+```bash
+# 1. 빌드
+cd sarimax_rs && maturin develop --release
+
+# 2. 종합 벤치마크 실행
+.venv/bin/python python_tests/bench_full_comparison.py
+
+# 3. Rust 내부 벤치마크
+cargo bench
+```
+
+## 5. 결과 판정 기준
+
+### PASS 조건 (모든 항목 충족 필요)
+
+- [x] **속도**: 10개 모델 전부 Rust >= 1.0x speedup (10/10 PASS)
+- [x] **정확도**: oracle loglike 오차 < 1e-6 (10/10 모델 PASS)
+- [ ] **정확도**: fitted params 오차 < 1e-2 (8/10 모델, ARIMA(2,1,2) + ARIMA(1,1,1) 실패)
+- [ ] **정확도**: forecast mean 오차 < 1e-4 (1/10 모델)
+- [x] **수렴**: 수렴률 >= 90% (10/10 모델 PASS)
+- [x] **배치**: 50-series batch speedup >= 3.0x (66.2x, PASS)
+
+---
+
+## 6. 벤치마크 결과 이력
+
+### 6.1 초기 결과 (2026-02-22, 옵티마이저 개선 전)
+
+#### 환경
+- Platform: macOS-15.1-arm64 (Apple Silicon)
+- Python: 3.14.3
+- sarimax_rs: 0.1.0
+- Config: method=lbfgsb (multi-start), maxiter=200
+
+#### 속도
+
+| 모델 | Rust (ms) | SM (ms) | Speedup | RS iter | SM iter |
+|------|-----------|---------|---------|---------|---------|
+| AR(1) n=200 | 0.3 | 2.0 | 6.6x | 21 | 3 |
+| AR(2) n=300 | 3.7 | 3.2 | 0.9x | 45 | 3 |
+| MA(1) n=200 | 0.5 | 2.3 | 4.8x | 30 | 4 |
+| ARMA(1,1) n=300 | 3.8 | 4.3 | 1.1x | 50 | 5 |
+| ARMA(2,2) n=400 | 32.2 | 13.6 | 0.4x | 67 | 11 |
+| ARIMA(1,1,1) n=300 | 8.5 | 7.1 | 0.8x | 43 | 11 |
+| ARIMA(2,1,2) n=400 | 67.2 | 66.3 | 1.0x | 87 | 36 |
+| SARIMA(1,0,0)(1,0,0,4) | 22.3 | 5.2 | 0.2x | 45 | 7 |
+| SARIMA(1,1,1)(1,1,1,12) | 2136.5 | 223.0 | 0.1x | 97 | 21 |
+| SARIMA(2,1,1)(1,1,1,12) | 1155.6 | 739.5 | 0.6x | 140 | 34 |
+
+**판정**: **FAIL** — 3/10 모델만 확실히 빠름
+
+#### 근본 원인
+- Multi-start 전략이 5~15개의 optimization run 수행 (불필요한 반복)
+- L-BFGS-B 파라미터가 scipy 기본값과 상이 (m=7, pgtol=1e-7)
+- AR/MA 파라미터에 불필요한 bounds [-20, 20] 적용 (L-BFGS-B Cauchy point 비효율)
+
+---
+
+### 6.2 최종 결과 (2026-02-22, 옵티마이저 개선 후)
+
+#### 환경
+- Platform: macOS-15.1-arm64 (Apple Silicon)
+- Python: 3.14.3
+- sarimax_rs: 0.1.0
+- Config: method=lbfgsb (single-run), maxiter=200
+
+#### 적용된 개선사항
+1. **Multi-start 제거**: 기본 `"lbfgsb"` 메서드를 단일 L-BFGS-B 실행으로 변경 (multi-start는 `"lbfgsb-multi"`로 분리)
+2. **L-BFGS-B 파라미터 정렬**: scipy 기본값과 동일하게 조정 (m=10, factr=1e7, pgtol=1e-5)
+3. **Fused function+gradient**: StateSpace 재생성 없이 분석적 그래디언트와 함수값을 동시 계산
+4. **Unbounded 파라미터**: enforce_stationarity/invertibility=true일 때 AR/MA bounds 제거 (Monahan/Jones 변환이 제약을 처리하므로 L-BFGS-B bounds 불필요)
+
+#### 속도
+
+| 모델 | Rust (ms) | SM (ms) | Speedup | RS iter | SM iter |
+|------|-----------|---------|---------|---------|---------|
+| AR(1) n=200 | **0.4** | 3.1 | **8.5x** | 21 | 3 |
+| AR(2) n=300 | **1.2** | 5.0 | **4.0x** | 17 | 3 |
+| MA(1) n=200 | **0.4** | 3.6 | **8.2x** | 16 | 4 |
+| ARMA(1,1) n=300 | **0.7** | 6.4 | **9.1x** | 10 | 5 |
+| ARMA(2,2) n=400 | **10.6** | 20.4 | **1.9x** | 40 | 11 |
+| ARIMA(1,1,1) n=300 | **3.0** | 10.6 | **3.6x** | 23 | 11 |
+| ARIMA(2,1,2) n=400 | **12.5** | 101.3 | **8.1x** | 27 | 36 |
+| SARIMA(1,0,0)(1,0,0,4) | **6.7** | 8.1 | **1.2x** | 25 | 7 |
+| SARIMA(1,1,1)(1,1,1,12) | **206.6** | 337.0 | **1.6x** | 20 | 21 |
+| SARIMA(2,1,1)(1,1,1,12) | **1085.2** | 1131.8 | **1.0x** | 49 | 34 |
+
+**판정**: **PASS** — 10/10 모델 전부 Rust >= 1.0x speedup
+
+#### 정확도
+
+| 모델 | Oracle LL | Param err | LL err | FC mean | FC CI |
+|------|-----------|-----------|--------|---------|-------|
+| AR(1) | 5.7e-14 | 6.9e-05 | 9.1e-07 | 2.8e-05 | 6.2e-03 |
+| AR(2) | 1.1e-13 | 1.6e-03 | 5.9e-04 | 5.6e-04 | 1.0e-02 |
+| MA(1) | 2.6e-10 | 3.8e-03 | 1.5e-03 | 3.5e-03 | 1.4e-02 |
+| ARMA(1,1) | 1.3e-10 | 1.3e-03 | 9.2e-04 | 2.1e-03 | 1.3e-02 |
+| ARMA(2,2) | 1.8e-10 | 2.7e-03 | 3.4e-04 | 1.6e-03 | 4.5e-03 |
+| ARIMA(1,1,1) | 5.0e-11 | 2.7e-02 | 1.1e-03 | 6.5e-03 | 1.3e-02 |
+| **ARIMA(2,1,2)** | 4.0e-08 | **1.5e+00** | **4.4e-01** | 1.5e-01 | 6.3e-01 |
+| SARIMA(1,0,0)(1,0,0,4) | 5.7e-14 | 2.9e-03 | 1.4e-03 | 4.0e-03 | 3.2e-02 |
+| SARIMA(1,1,1)(1,1,1,12) | 2.6e-10 | 3.5e-03 | 1.1e-03 | 3.3e-02 | 7.1e-02 |
+| SARIMA(2,1,1)(1,1,1,12) | 1.2e-09 | 1.8e-03 | 6.8e-04 | 2.3e-02 | 5.4e-02 |
+
+#### 배치
+
+| 시나리오 | Rust (ms) | SM (ms) | Speedup | 수렴 |
+|----------|-----------|---------|---------|------|
+| AR(1) 50x n=200 | **2.4** | 155.6 | **66.2x** | 50/50 |
+
+#### 종합 판정
+
+| 항목 | 결과 | 상세 |
+|------|------|------|
+| 속도 (전 모델 우위) | **PASS** | 10/10 (1.0x ~ 9.1x) |
+| Oracle loglike < 1e-6 | **PASS** | 10/10 |
+| Param error < 1e-2 | **FAIL** | 8/10 (ARIMA(2,1,2), ARIMA(1,1,1) 실패) |
+| Forecast mean < 1e-4 | **FAIL** | 1/10 |
+| 수렴률 >= 90% | **PASS** | 10/10 |
+| Batch speedup >= 3x | **PASS** | 66.2x |
+
+---
+
+## 7. 잔존 이슈 분석
+
+### 7.1 ARIMA(2,1,2) 파라미터 수렴 문제
+- param error 1.5 → statsmodels와 다른 국소 최적점에 수렴
+- oracle loglike는 4.0e-08으로 정확 → Kalman filter 자체는 정상
+- 다차원 MA가 포함된 모델에서 likelihood surface의 다중 최적점 문제
+
+### 7.2 ARIMA(1,1,1) 파라미터 오차 (경미)
+- param error 2.7e-02 → 허용 범위(1e-2) 초과이나 borderline
+- LL error 1.1e-03으로 작음 → 유사한 loglike를 가진 근접 파라미터에 수렴
+
+### 7.3 Forecast 오차
+- forecast mean 오차는 대부분 fitted params 차이에서 기인
+- oracle loglike가 1e-14 수준이므로 Kalman filter/forecast 로직 자체는 정확
+- 향후 param 정확도 향상 시 자동 개선 예상
+
+---
+
+## 8. 개선 로드맵
+
+### Phase 1: 옵티마이저 효율 개선 — ✅ 완료
+- [x] Multi-start 전략 제거 → 단일 L-BFGS-B 실행
+- [x] L-BFGS-B 파라미터 scipy 기본값 정렬 (m=10, factr=1e7, pgtol=1e-5)
+- [x] Fused function+gradient (분석적 그래디언트 + StateSpace 재사용)
+- [x] Unbounded 파라미터 (enforce시 bounds 제거)
+
+### Phase 2: 단일 모델 속도 검증 — ✅ 완료
+- [x] 옵티마이저 개선 후 재벤치마크
+- [x] 목표 달성: 모든 모델에서 speedup >= 1.0x
+
+### Phase 3: Forecast 정확도 향상 (향후)
+- [ ] ARIMA(2,1,2) 다중 국소 최적점 탐색 전략
+- [ ] ARIMA(1,1,1) param error 경감 (2.7e-02 → 1e-02 이하)
+- [ ] Forecast는 param이 동일하면 정확함 (oracle loglike PASS로 검증됨)
+
+## 9. 데이터 생성 방식
+
+모든 벤치마크 데이터는 `np.random.seed(42)` 기반 결정적(deterministic) 생성:
+
+- AR 데이터: `y[t] = phi * y[t-1] + noise`
+- ARIMA 데이터: `cumsum(noise)` + AR/MA 구조
+- SARIMA 데이터: `y[t] = phi * y[t-1] + Phi * y[t-s] + noise`
+
+이를 통해 재현 가능한 벤치마크 결과를 보장한다.