Analisis Model Sequence-to-Sequence Berbasis Gated Recurrent Unit (GRU) dengan Embedding dan Multi-Task Learning untuk Estimasi Cadangan Klaim Asuransi

ARNAYA INGGID NIRMALA, Prof. Dr. Drs. Gunardi, M.Si.

2026 | Skripsi | S1 ILMU AKTUARIA

Cadangan klaim adalah komponen penting kesehatan keuangan asuransi karena memengaruhi laporan keuangan dan solvabilitas, sehingga estimasinya perlu akurat agar tidak terjadi kondisi under-reserving  atau over-reserving. Chain Ladder mengandalkan faktor perkembangan agregat sehingga belum mempelajari ketergantungan antarperiode secara langsung, mendorong penelitian ini menggunakan sequence-to-sequence berbasis GRU mengacu kerangka DeepTriangle}sebagai alternatif. Penelitian ini menganalisis pengaruh company embedding dan multi-task learning yang memanfaatkan informasi klaim outstanding (OS) sebagai input tambahan dan target auxiliary, baik secara terpisah maupun kombinasi, terhadap performa estimasi, serta membandingkannya dengan Chain Ladder pada Private Passenger Auto (short-tail) dan Workers Compensation (long-tail), menggunakan data cadangan klaim berbentuk agregat dengan 50 perusahaan per lini.

Dalam penelitian ini, dilakukan pemodelan empat arsitektur model yang dievaluasi dengan MAPE dan RMSPE serta diuji signifikansinya terhadap Chain Ladder melalui Wilcoxon Signed-Rank. Seluruh model GRU terbukti lebih akurat daripada Chain Ladder, meski kontribusi embedding dan multi-task learning terhadap performa terbaik berbeda antara kedua lini bisnis. Analisis bias juga menunjukkan sebagian besar model cenderung over-reserving dengan tingkat ketidakpastian yang tergolong stabil. Hasil penelitian menunjukkan kompleksitas arsitektur tidak selalu meningkatkan akurasi, melainkan bergantung karakteristik lini bisnis, sehingga desain model perlu disesuaikan dengan data untuk estimasi cadangan klaim yang lebih andal.

Claim reserves are critical to an insurer's financial health, as they directly shape financial statements and solvency, so accurate estimation is essential to avoid under- and over-reserving. The Chain Ladder method relies on aggregate development factors and does not directly capture period-to-period dependencies in claims data, motivating this study to adopt a GRU-based sequence-to-sequence approach, following the DeepTriangle framework, as an alternative. The study examines how company embedding and multi-task learning affect performance, individually and combined, by leveraging outstanding claim (OS) information as both an additional input channel and an auxiliary output target, benchmarking the results against the Chain Ladder method on Private Passenger Auto (short-tail) and Workers Compensation (long-tail) using CAS data covering 50 companies per line.

In this study, four model architectures were developed and evaluated using MAPE and RMSPE, with their significance against Chain Ladder tested via the Wilcoxon Signed-Rank test. All GRU configurations proved more accurate than Chain Ladder, though the contribution of embedding and multi-task learning to the best model differed between the two lines. Bias analysis showed most models tended toward over-reserving, with stable uncertainty. These findings indicate that architectural complexity does not universally improve accuracy but depends on business line characteristics, so model design should be tailored to the underlying data.

Kata Kunci : Cadangan Klaim, Asuransi, Gated Recurrent Unit, Sequence-to-Sequence, DeepTriangle, Company Embedding, Multi-Task Learning, Chain Ladder.