每日沽空成交量能否預測沽空未平倉量？以 660,246 個結算窗口測試三排名次公式、Ridge 與小型 Transformer

FINRA publishes short volume every day and short interest only twice a month, so people treat the daily "short volume ratio" as a live read on short positions. They aren't the same thing: short volume is a flow of executions (a market maker can sell short to fill a buy order and cover an hour later), short interest is the position still open at settlement. I tested how much the daily flow actually tells you about the next short-interest report, on 660,246 settlement windows across 6,959 stocks from January 2020 to July 2026.

The target is the percentage change in reported short interest over the same settlement window. Over the whole 2020 to 2026 panel, the plain ratio change reaches +0.229.

Five findings

1. The signal is real but concurrent. The ratio change correlates +0.225 with the short-interest change in its own window, -0.118 with the previous one and +0.012 with the next. It's a nowcast of the report that hasn't been published yet, not a forecast of the one after.

2. 50% isn't the line. Short interest rose in 37.4% of windows with a ratio of 35% or lower and 63.2% above 60%, but the 50/50 crossover sits near the market's average ratio (46.4%), not at 50%. A market-relative ratio was positive on all 156 settlement dates.

3. Three ranks beat 32 hand-tuned formulas. The best transparent score averages three within-date percentile ranks: the ratio change, the ratio minus the market average, and excess short shares (actual short volume minus what the stock's recent ratio implies at today's volume) divided by prior short interest. It improved on the baseline in every period: 2020 to 2023, the 2024 validation year and the 2025 to 2026 test.

4. Heavy volume makes the ratio more informative. On the stock's hottest-volume fifth of windows, the gap in "short interest rose" between the highest and lowest ratio quintiles was almost 28 points, against 21 on the quietest. Volume relative to the stock's own history mattered (interaction +0.037, t = 13.0); raw volume, which mostly measures company size, barely did (+0.008).

5. It isn't a trading signal. The next-window return correlations were about zero (-0.005 for the ratio change). Within the hottest-volume fifth, a high market-relative ratio correlated -0.036 with the next two weeks' return (t = -7.6): detectable, but small and concentrated where trading costs are highest.

The model

• Each sequence position is one FINRA trading day with 21 features (short ratio, exempt and non-exempt flow, volume against 21- and 63-day history, flow against prior short interest, days to cover, missingness masks), plus a branch of 14 window aggregates with 14 availability masks.
• Days are projected to 32 dimensions, then two causal self-attention layers (four heads, 96-unit feed-forward). Heads output the median log change, P10 and P90 bounds, and an auxiliary rank score.
• Huber loss (delta 0.15) on the log change, plus 0.2-weighted pinball loss for the interval and MSE on the within-date rank. AdamW, batch 512, lr 0.0005, up to eight epochs, early stopping on 2024.
• Trained through 2023, selected on 2024, tested on 2025 to July 2026. Seeds 17, 29 and 43 scored +0.408, +0.405 and +0.404; the ensemble +0.414, with 63.4% direction accuracy and an 82.4% P10 to P90 coverage.
• Weights, ONNX graph and the input contract are on Hugging Face under Apache-2.0.

The traps I hit

• Final files aren't point-in-time. FINRA's history in my store keeps the latest version of each file, all imported on one day, so this is a retrospective test on final data. A true forward test has to archive each prediction before the official number comes out.
• The test period was looked at. I'd inspected 2025 to 2026 during the formula search, so +0.414 is a strong retrospective result, not a promised live correlation.
• No current fundamentals. Today's market cap and shares outstanding would leak the future into old windows, so they're excluded from every model.
• Rank within the date. Every rank and quintile is formed inside one settlement date; ranking across dates mixes market regimes.
• Follow the settlement calendar. Windows run from one settlement date to the next, including the May 2024 switch from T+2 to T+1.
• Short-exempt volume doesn't help. It's about 1.33% of short volume, and subtracting it slightly lowers the correlation. Using only the last three days of a window lowers it a lot: the information is spread across the window.

Try it on one stock

With the Equibles MCP server connected to Claude or ChatGPT, a prompt like "Using Equibles, compare GameStop's daily short volume ratio over the current settlement window with the market average, and show the last six short-interest reports" gives you the inputs for the three-rank score on one name.

The full study has the formulas with a worked example, the quintile tables, the ridge benchmark, the robustness checks and the prospective validation protocol: Does Daily Short Volume Predict Short Interest? Evidence From 660,000 Settlement Windows.