法拍屋評價模型與投資策略之研究 ——台北市之實證分析

Abstract

本研究以 2024 年 1 月至 2026 年 2 月間台北市地方法院強制執行拍賣案件（有效拍次 734 筆，Layer A 257 筆，估值前排除凶宅與有建無地後主迴歸樣本 251 筆）為研究對象，建構一套以「兩把尺」為核心之系統性法拍屋估值與投資決策框架：第一把尺為 RSD 自動估值模型（提供市場價值基準），第二把尺為查封折扣調整系統（FAS，以法拍特定因子修正估值），並整合為 FIDS 法拍投資決策系統。 估值模型方面，本研究提出反對稱分解設計 AVM（RSD AVM），以建物特徵對稱乘數（BD）與 K 近鄰地段基準價（BE）構成雙層估值結構，並將車位持分獨立估價，以 avm_final 為正本估值。以台北市 20,176 筆實價登錄資料執行留一法全樣本回測（LOOCV），達 MAPE = 9.62%、MdAPE = 6.80%，中位帶號誤差近 0，無系統性偏誤。 查封折扣調整方面，本研究建立 FAS，自法院查封筆錄以 NLP（採「肯定句＋排除查詢樣板＋排除否定句」三層清洗）萃取事前可觀測之外生法拍因子，全部以二元虛擬變數納入 OLS 迴歸（應變數為折溢比值 成交價／avm_final；採 White 異質變異穩健標準誤）。為使迴歸樣本與可投資母體一致，凶宅（2 筆）與有建無地（4 筆，折逾 78% 之高槓桿極端案）一律於估值前排除，主模型以 251 筆估計。主模型採無截距設定，使每一係數直接讀作「相對市場完美點交（ratio=1，即正常市場乾淨交付）之水準」，毋須解釋截距。沿點交狀態呈一道折價階梯：法拍完全點交即折 16.9%（t = −7.87，驗證『法拍點交 ≠ 市場點交』）、部份點交 17.3%、不點交 18.6%、現況點交 24.7%（中心化 R² = 0.139）；輔以查封筆錄原文,四類含意分明（部份點交僅邊際附屬不點交故近完全點交,現況點交折最深，推測與標的核心較複雜相關）。其他外生因子中，遺產分割（+0.158***）、含停車位（+0.094***）達顯著、未保存登記（+0.068*）邊際；傳統認知之違建、占用、可修復瑕疵等多不顯著。偏低之解釋力屬合理——應變數為市場價值（avm_final）已知後之殘差折價比值、而非房價本身，故 R² 僅衡量「殘差折價中可由公開文字解釋之比例」；其餘變異或反映投資人私有資訊等不可觀測因素（即法拍市場資訊高度不對稱，推測）。而市場拒絕深度、流標次數、投標人數等事後實現之內生變數，因倒果為因且事前不可觀測，一律排除於估值模型之外，此為本研究計量設計之關鍵立場。 雙折扣與投資應用方面，本研究以 FAS-AVM（avm_final 乘上「1＋各因子係數之和」）還原物件之風險調整公允價值，並對比「未調整之市場 AVM 折扣」與「FAS 調整後之真實折扣」。於雙折扣分析池（自 Layer A 排除 2 筆確認凶宅與 4 筆有建無地後之 251 筆）中，FAS 真實折扣中位數達 18.2%（深於未調整之 11.5%），並辨識出偽折扣（投資陷阱）1 筆與隱藏折扣 23 筆。據此提出 FIDS 六階篩選管線（首步排除絕對不碰之特殊件）與 AA 推薦原則，最終對待標物件產出可直接比對法院底價之「出價上限」投標建議清單。符合投資原則之標的數量充足且分級具鑑別力（FIDS 完整過濾後合格投資池 233 筆中，過損益平衡者 181 筆，AA 強推 37 筆）。壓力測試顯示，AA 強推池於高度壓力（持有兩年、修繕 7.5%、售價折讓 5%、交易成本 5%、資金成本 8%）下仍有 97% 維持正報酬、ROI 中位 12.4%，而全可投資池則降至 −9.1%（僅 29% 正報酬），反向確立 AA 安全邊際之必要性。顯示問題不在尋找案源，而在精準過濾偽折扣、辨識隱藏折扣。跨樣本穩健性檢定（Layer B，N=118）顯示主要外生因子之係數方向與 Layer A 一致，支持主模型之穩健性。 本研究建立之兩把尺框架可供後續跨城市比較、拍定後追蹤研究及實務投資系統開發參考。

This study constructs a "two-ruler" framework for systematic foreclosure valuation and investment decision-making, using 734 valid foreclosure auction cases from the Taipei District Court during January 2024 to February 2026 (Layer A: 257; main regression sample after pre-estimation exclusions: 251). The first ruler is the RSD Automated Valuation Model (market value benchmark); the second is the Foreclosure Adjustment System (FAS), which corrects the valuation using foreclosure-specific factors. The two are integrated into the FIDS (Foreclosure Investment Decision System). Regarding the valuation model, this study proposes the Reverse-Symmetric Design AVM (RSD AVM), combining building characteristic symmetric multipliers (BD) with a K-nearest-neighbor base estimate (BE), and valuing parking shares independently to obtain the canonical estimate avm_final. Leave-One-Out Cross-Validation (LOOCV) on 20,176 real-price-registration records achieves MAPE = 9.62% and MdAPE = 6.80%, with a median signed error near zero, indicating no systematic bias. Regarding the Foreclosure Adjustment System (FAS), this study extracts ex-ante observable exogenous foreclosure factors from court attachment records via NLP (with a three-layer cleaning rule: affirmative statements + exclusion of due-diligence boilerplate + exclusion of negations), all entered as binary dummies in an OLS regression (dependent variable: the price ratio sold price / avm_final, White heteroskedasticity-robust standard errors). To align the regression sample with the investable population, stigmatized "haunted" properties (2 cases) and building-without-land cases (4 cases, −78%, high-leverage extremes) are excluded before estimation, so the main model is fit on 251 cases. The main model uses a no-intercept specification, so every coefficient reads directly as a level "relative to a market-perfect handover (ratio = 1, i.e., a normal clean-delivery market sale)," with no constant to explain. Along the handover status it forms a discount ladder: even a fully point-delivered foreclosure already discounts 16.9% (t = −7.87, validating "foreclosure handover ≠ market handover"); partial handover 17.3%; no handover 18.6%; as-is handover 24.7% (centered R² = 0.139); court-record text clarifies the four meanings (partial handover withholds only peripheral shares, hence ≈ full; as-is handover is likely associated with a more complex core, presumably the deepest). Among other exogenous factors, inheritance/estate partition (+0.158***) and parking (+0.094***) are significant and unregistered building (+0.068*) marginal; conventionally assumed factors—illegal construction, occupation, repairable defects—are mostly insignificant. The low explanatory power is to be expected, since the dependent variable is the residual discount ratio after market value (avm_final) is already accounted for—not the house price itself; the R² thus measures only the share of the residual discount explainable by public textual signals, with the remainder likely reflecting unobservable factors such as investors' private information (i.e., the foreclosure market's high information asymmetry). Furthermore, ex-post realized endogenous variables (market rejection depth, failed-bid count, number of bidders) are excluded from the valuation model, as they are reverse-causal and unobservable at bidding time—a key methodological stance of this study. Regarding the dual-discount analysis and investment application, this study uses FAS-AVM (avm_final multiplied by "1 + the sum of factor coefficients") to recover the risk-adjusted fair value, and contrasts the "unadjusted market-AVM discount" with the "FAS-adjusted true discount." Within the dual-discount analysis pool (251 cases, after excluding 2 confirmed stigmatized "haunted" properties and 4 building-without-land cases from Layer A), the median FAS true discount reaches 18.2% (deeper than the unadjusted 11.5%), identifying 1 spurious discount (investment trap) and 23 hidden discounts. Based on this, the study proposes the FIDS six-stage filtering pipeline (with exclusion of absolutely-avoid special cases as the first step) and AA recommendation rules, ultimately producing a "bid-ceiling" recommendation list directly comparable to court reserve prices. The number of principle-compliant targets is ample with discriminating grades (within the fully FIDS-filtered pool of 233, 181 are above break-even and 37 are AA strong-buy). Stress testing shows that under a severe scenario (two-year holding, 7.5% renovation, 5% resale haircut, 5% transaction cost, and 8% capital cost) the AA pool still yields 97% positive returns and a median ROI of 12.4%, whereas the full investable pool falls to −9.1% (only 29% positive)—reverse-validating the necessity of the AA safety margin. This indicates that the challenge is not sourcing cases but precisely filtering spurious discounts and identifying hidden ones. A cross-sample robustness check (Layer B, N=118) shows that the signs of the main exogenous factors are consistent with Layer A, supporting the robustness of the main model. The two-ruler framework established in this study may serve as a reference for future cross-city comparative studies, post-auction longitudinal tracking, and practical investment system development.