SOFR / Treasury Spread Modeling

Project Summary

This repository documents a full modeling progression for yield estimation on bond data using a benchmark-plus-spread framework.

It includes:

• early direct regression baselines,
• benchmark-aware spread models,
• issuer-level and bond-level decompositions,
• research upper-bound modeling,
• final no-leakage model.

The final clean result was a no-leakage bond-plus-issuer benchmark model with mean error around 5.8 bps and median around 3.7 bps.

Core Formula

The final family of models used:

Y_i = B(T_i) + alpha_issuer(i) + beta_bond(i) + gamma^T x_i

where:

• Y_i = reported yield
• B(T_i) = benchmark Treasury curve at maturity T_i
• alpha_issuer(i) = issuer-level spread effect
• beta_bond(i) = bond-level spread effect
• x_i = residual features such as coupon, discount, and maturity

Model Progression

V6 — Standard baseline

Benchmark-plus-spread style model with issuer / feature structure.

• Good baseline
• Around 12 bps error
• Easy to explain

V8 — Cleaner structured baseline

Uses:

• smooth benchmark curve,
• issuer effect,
• residual regression. Still around 12 bps, but more defensible and better organized.

V9 — Shrinkage test

Applied shrinkage to issuer effects.

• Error worsened
• Important finding: issuer signal was already strong

V10 — Research upper bound

Adds bond-specific adjustments in-sample.

• Around 4–5 bps
• Too optimistic for production
• Useful as an upper bound

V11 — Final clean model

Uses leave-one-out issuer and bond effects to reduce leakage.

• Around 5.8 bps mean
• Around 3.7 bps median
• Best defensible model

Repository Structure

• v8_model.py — clean benchmark + issuer model
• v10_model.py — in-sample bond-level upper-bound model
• v11_model.py — final no-leakage research-grade model
• model_progression.md — paper-style methodology and model evolution

Important Positioning

• V8 = production-style baseline
• V10 = research upper bound
• V11 = best clean model