納入EFFR與FFF之SOFR利率預測：Vasicek和LSTM模型預測之比較

Abstract

SOFR利率逐步取代USD LIBOR成為主要金融商品訂價參考指標。有鑒於轉換參考利率對於金融市場和實體經濟都會產生重大影響，因此有必要製作一有效且可靠的SOFR利率預測模型。本研究以Vasicek和LSTM模型驗證多變量模型預測利率的優勢。首先，將聯邦基金利率期貨和聯邦資金有效利率分別納入Vasicek和LSTM模型，實證結果指出納入上述兩個因子可以強化模型捕捉貨幣政策變動的能力，減少原本模型在貨幣政策變動後預測值短期內偏離實際值的幅度。此結果亦符合經濟直覺，即模型考量更多有效資訊後表現更佳，此結果在Vasicek和LSTM模型上都得到驗證。其次，利用LSTM模型靈活度高、容易納入更多預測變數之優點，再將總體經濟因子加入已經含有前述提及之聯邦基金利率期貨、聯邦資金有效利率和SOFR歷史資料的LSTM模型之中，發現結合貨幣政策和總體經濟因子以及SOFR歷史資料的LSTM模型，於降息期間預測表現優於僅納入SOFR歷史資料和貨幣政策相關因子的LSTM模型。不過LSTM模型於升息期間納入總體因子後，預測表現之改善幅度並未如同降息期間一般顯著，此結果代表於不同貨幣政策時期，納入總體因子所能減少的模型誤差幅度並不相同。

As USD LIBOR to SOFR transition has started, it is crucial to construct a reliable model for forecasting the SOFR rate. This study justifies the advantages of using a multivariate model to forecast interest rates. Initially, we employ both Vasicek and LSTM models, incorporating the federal funds rate futures and effective federal funds rate as input factors. The empirical results indicate that including these two factors improves the model`s ability to reduce the deviation between predicted and actual values after monetary policy changes. This result aligns with economic intuition, as including more relevant information enhances the model`s performance. Both Vasicek and LSTM models justify the advantages of using a multivariate model. Moreover, the LSTM model performs better after incorporating macroeconomic factors alongside the aforementioned federal funds rate futures, effective federal funds rate, and historical SOFR data, especially during rate-cutting periods. That is, the improvement of performances after incorporating more relevant macroeconomic factors into the LSTM model during rate-hiking periods is less significant than during rate-cutting periods. This result suggests that the reduction in model error achieved by including macroeconomic factors varies depending on the different monetary policy periods.