Predictive Models · Direction

WaveCast-D 1.0.0

Next-day direction prediction for US equity special stocks. A calibrated Up/Down call and conviction score for every active special, every trading day before 8:00 AM EST.

70.04%
Direction Accuracy
419K
OOS Predictions
+10pp
vs Naive Baseline
84.5%
Accuracy at Q5 Conf.
All metrics out-of-sample · 419,505 predictions · 122 trading days · July 2025 – January 2026

What It Does

A binary direction call — with a conviction score that tells you how much to trust it.

Each morning WaveCast-D scores every active US equity special — any stock with an intrinsic borrow rate above 20 basis points — producing a predicted direction for the next trading day and a confidence score calibrated to realized accuracy. The score is not a raw probability; it is a monotonically calibrated conviction measure: higher score, higher realized accuracy.

A gradient-boosted classifier trained on rate history, market microstructure signals, cross-sectional universe dynamics, and demand-side positioning data. Walk-forward validated across 20 time-ordered folds and four distinct OBFR rate regimes — low, hiking, high, and easing — with no look-ahead and no data leakage at any step.

Daily Output Schema
What your team receives each morning, per ISIN.
Field Description
Date Scoring date
ISIN Security identifier
PredictedDirection 1 = Up, 0 = Down
ProbabilityUp Raw model probability (0–1)
DirectionConfidence Calibrated conviction score (0–100)
IntrinsicRateAvg_T1 Current day’s actual rate (context)
Delivery
AWS S3 · daily before 8:00 AM EST · CSV + Parquet

The Research Evidence

Three views into the model.

Accuracy vs naive baseline, confidence calibration, and per-tier consistency — three views demonstrating genuine predictive skill across the full special universe.

Evidence 01

+10 Percentage Points Above the Naive Baseline

The naive baseline — always predict Down, exploiting the structural mean-reversion tendency in special rates — achieves roughly 60% accuracy. WaveCast-D achieves 70.04% across 419,505 out-of-sample predictions. That 10pp gap represents genuine predictive skill over an informed baseline.

Against the persistence baseline (yesterday’s direction = tomorrow’s), WaveCast-D is +33pp. Special rates mean-revert structurally — persistence is anti-predictive in this market.

Direction Accuracy vs Baselines
419,505 out-of-sample predictions · July 2025 – January 2026
Realized Accuracy by Confidence Quintile
DirectionConfidence (0–100) · monotonically calibrated
Evidence 02

84.5% When the Model Says It’s Confident

The DirectionConfidence score (0–100) is monotonically calibrated to realized accuracy. Q5 — the highest-confidence quintile — achieves 84.5% realized accuracy. Q1 — lowest confidence — achieves 59.6%, still above the naive baseline.

This calibration is the operational signal: filter to Q4+ to concentrate on the highest-accuracy subpopulation, or use the raw score as a continuous weight for portfolio construction.

Evidence 03

Consistent Across All Seven Rate Tiers

Direction accuracy holds between 65.66% and 69.88% across all seven rate tiers — from Barely Warm shallow specials to Very Extreme deep squeezes. The model does not derive its headline number from one well-performing subgroup.

Accuracy is slightly lower in the deepest tiers (Very Hot, Extreme, Very Extreme) where rate dynamics are more squeeze-driven and less predictable. This is expected and disclosed.

Direction Accuracy by Depth Tier
Walk-forward validation · all 20 folds · 6 years
Methodology

Walk-forward cross-validation across 20 time-ordered folds, September 2020 – January 2026. No look-ahead at any step. Training windows never include validation or test data. Model retrained monthly; weekly drift monitoring triggers unscheduled retraining when rolling 5-day accuracy drops more than 5 percentage points below the 63-day baseline. Headline accuracy (70.04%) is measured on the 122-trading-day validation window, July 2025 – January 2026 (419,505 predictions). Walk-forward accuracy across all 20 folds: 70.55% (3,643,852 predictions). Held-out test set accuracy: 68.59%. Live production accuracy (59 scored dates, 203,084 rows): 67.98%.

Validated Across Every Horizon

The same model, tested four different ways.

70.55%
Walk-Forward
20 Folds, 3.6M rows
70.04%
Validation Window
419K predictions
68.59%
Held-Out Test
Promotion Benchmark
67.98%
Live Production
59 dates, 203K rows

How Clients Use It

Three institutional applications.

WaveCast-D is designed for securities lending desks and agency lenders managing daily inventory deployment decisions.

Inventory Deployment

Push/Hold Decisions

Use DirectionConfidence to calibrate same-day inventory deployment. High-confidence Up predictions support pushing available stock into the market today; high-confidence Down predictions support holding for rate recovery. Pair with WaveCast-R for magnitude context.

Rate Negotiation

Counterparty Positioning

Enter rate conversations with a directional view. A high-confidence Down prediction on a hot name is information a counterparty may not have. Use it to negotiate term locks before the market moves, or to time recall decisions on existing open loans.

Portfolio Scoring

Universe Prioritization

Score the full ~4,000 ISIN universe each morning, then filter to high-conviction Up names in the deepest rate tiers. Concentrate trader attention and client outreach on the names most likely to move in your favour before the close.

Request beta access to WaveCast-D.

WaveCast is an invite-only beta. We show you WaveCast predictions against your inventory first. No generic demos or hard sales pitches. Share your name, firm, and role and we’ll be in touch.

Explore the WaveCast Suite

THE INFORMATION CONTAINED HEREIN IS PROPRIETARY TO TIDAL MARKETS LLC AND DESIGNATED THIRD PARTIES. THESE ILLUSTRATIONS MAY NOT BE COPIED, EDITED, OR REDISTRIBUTED WITHOUT THE EXPRESS WRITTEN CONSENT OF TIDAL MARKETS LLC. ANY INFORMATION CONTAINED HEREIN DOES NOT CONSTITUTE OR IMPLY INVESTMENT ADVICE BY TIDAL MARKETS LLC OR ANY REPRESENTATIVES THEREOF. TIDAL MARKETS LLC DOES NOT GUARANTEE THE QUALITY, ACCURACY, OR COMPLETENESS OF THE INFORMATION CONTAINED HEREIN. ALL MATERIALS PROVIDED ARE FOR INFORMATIONAL PURPOSES ONLY AND IS NOT INTENDED FOR TRADING PURPOSES, NOR SHOULD BE CONSTRUED OR INTERPRETED AS A COURSE OF ACTION. TIDAL MARKETS LLC IS NOT RESPONSIBLE OR LIABLE IN ANY FORM TO ANY READER, FIRM OR CORPORATION, FOR ANY DAMAGES OR LOSSES ARISING FROM ANY USE OF THE INFORMATION CONTAINED HEREIN. PAST PERFORMANCE IS NOT INDICATIVE OF FUTURE RESULTS.