本研究採用中央氣象署1980至2015年地震目錄，評估最強地震圓弧雙交叉（SDICAE）對臺灣短期強震之預測表現，以震源深度20公里以內之事件，分別設定短週期（dY＝2）與長週期（dY＝6）進行搜尋；經隔日確認取得600筆交叉，型態分類後保留413筆完全交叉，去重複後建立361筆分析母體，實測天數以代碼8888標示之23筆紀錄，依命中定義判定為未命中並納入分母。 分析依半徑比、端弧比、主軸角、最大分岔比及兩端分岔對稱比逐級篩選，並以時空完全隨機之蒙地卡羅模擬估計背景命中率，配合單尾精確二項檢定、Clopper–Pearson信賴下限、技巧比值及多重檢定校正評估，命中定義為位於預測區內、芮氏規模5.7以上、震源深度70公里以內，且實測天數在120天以內之強震，確認形成日當日發生者亦計入命中。 母體命中率為30.5%（110／361），單尾95%信賴下限為26.5%，相對23.28%背景值之技巧比值為1.31，單尾二項檢定p＝0.0010，校正後仍顯著，五項幾何門檻逐級施加後，合計命中率降至24.0%；雙凸交叉由29.2%（42／144）提高至50.0%（5／10），技巧比值由1.25增至2.15；凹凸交叉則由31.3%（68／217）降至17.5%（7／40），技巧比值由1.35減至0.75；四組週期型態分列時，雙凸交叉於dY＝2與dY＝6分別由30.2%及27.1%升至37.5%及100.0%，凹凸交叉則由32.7%及28.4%降至16.0%及20.0%，兩種交叉型態對幾何篩選之反應方向相反。 dY＝2與dY＝6之母體命中率分別為31.7%及27.8%（p＝0.540），兩者具事件涵蓋互補性，事件命中率為28.3%（49／173），四組聯集涵蓋49起強震，所得結果仍屬回溯性證據，有待前瞻性資料驗證。;This study evaluated the performance of the Strongest Double Intersections of Circular Arcs of Earthquakes (SDICAE) for short-term forecasting of strong earthquakes in Taiwan， using the 1980–2015 earthquake catalog maintained by the Central Weather Administration. Events with focal depths of 20 km or less were examined over a short search period (dY = 2) and a long search period (dY = 6). Next-day confirmation identified 600 intersections. Following type classification， 413 complete intersections were retained， and deduplication produced a final analysis sample of 361 intersections. The observed-time field was coded as 8888 for 23 records; under the adopted hit definition， these records were classified as misses and included in the denominator. The intersections were progressively screened using five geometric criteria: radius ratio， end-arc ratio， major-axis angle， maximum branching ratio， and branching symmetry ratio. The background hit rate was estimated using a fully randomized spatiotemporal Monte Carlo simulation. Performance was evaluated using one-sided exact binomial tests， Clopper–Pearson one-sided confidence lower bounds， skill ratios， and correction for multiple testing. A hit was defined as an earthquake of magnitude 5.7 or greater and focal depth of 70 km or less that occurred within the prediction zone at 0 ≤ dt ≤ 120 days. Events occurring on the confirmation date were also counted as hits. The overall hit rate was 30.5% (110/361)， with a one-sided 95% confidence lower bound of 26.5%. Relative to the background hit rate of 23.28%， the skill ratio was 1.31， and the one-sided binomial test yielded p = 0.0010; the result remained significant after multiple-testing correction. Progressive application of the five geometric criteria reduced the overall hit rate to 24.0%. For double-convex intersections， the hit rate increased from 29.2% (42/144) to 50.0% (5/10)， with the skill ratio rising from 1.25 to 2.15. In contrast， the hit rate for concave-convex intersections decreased from 31.3% (68/217) to 17.5% (7/40)， with the skill ratio declining from 1.35 to 0.75. When the four search-period–type groups were considered separately， the hit rates for dY = 2 and dY = 6 double-convex intersections increased from 30.2% and 27.1% to 37.5% and 100.0%， respectively， whereas those for concave-convex intersections decreased from 32.7% and 28.4% to 16.0% and 20.0%. The two intersection types therefore showed opposite responses to geometric screening. At the population level， the hit rates for dY = 2 and dY = 6 were 31.7% and 27.8%， respectively， with no statistically significant difference between them (p = 0.540). Nevertheless， the two search periods were complementary in terms of event coverage. The event-level hit rate was 28.3% (49/173)， and the union of the four groups covered 49 distinct strong earthquakes. These findings are based on retrospective evidence and require validation using prospective data.
