DeriveAI is a ROUTUR, Inc. company.
The models are built and evaluated by DeriveDX Quantitative Labs, ROUTUR's research arm.
FTO-1 has a large temporally held-out evaluation: 325,463 candidate outcomes across 75,417 jobs, starting 2026-04-13. Its calibrated test ROC-AUC is 0.7965, average precision is 0.5747, and calibration ECE is 0.0229. Ranking the top candidate per job produced a 27.16% target-hit rate, compared with 24.28% for the strongest legacy ranker.
FTO-1.5 has a separate, policy-conditioned July research holdout. On the controlled fixed-policy comparison, it improved the top-5% result to 69.4% hit rate and +0.541R, versus FTO-1’s 64.5% and +0.468R. The FTO-1.5 artifact is still shadow-only and its results are diagnostic.
A forward FTO-1 scoring update is complete: 13,263 Aug. 6+ grade records now have versioned predictions from the frozen FTO-1 artifact. Its first forward reconstruction, across 4,772 eligible resolved candidates, has a raw-score ROC-AUC of 0.9971 and average precision of 0.9931. This is an early two-entry-day result, not a mature live-performance claim.
| Term | Meaning in this report |
|---|---|
| Target hit | Take profit reached before stop loss or expiry, under the model’s stated policy. It is not an overall profitable-trade or portfolio-return measure. |
| R | Realized return normalized by the policy’s risk basis. It excludes commissions, slippage, position sizing, capital constraints, and portfolio overlap. |
| Temporal holdout | A date-separated test period excluded from model fitting and calibration. It is historical out-of-sample evidence, not live-trading performance. |
| Forward monitoring | Predictions captured from a frozen artifact for recommendations created after its freeze. A forward metric is publishable only when the predictions were captured before outcome resolution and the cohort has matured. |
FTO-1 is a binary classifier for take profit before stop loss. Its training, validation and test dates are split chronologically; the test period begins on 2026-04-13 and contains 325,463 candidate outcomes after expiration overlap purging.
| Held-out metric | Result |
|---|---|
| Test candidates / jobs | 325,463 / 75,417 |
| Target base rate | 21.31% |
| ROC-AUC (calibrated) | 0.7965 |
| Average precision (calibrated) | 0.5747 |
| Brier score (calibrated) | 0.1285 |
| Log loss (calibrated) | 0.4083 |
| Calibration ECE, 15 bins | 0.0229 |
| FTO-1 top-1 target-hit rate | 27.16% |
| Best legacy top-1 target-hit rate | 24.28% |
| Absolute top-1 lift | +2.87 percentage points |
FTO-1.5 predicts whether a requested take-profit/stop-loss policy will reach take profit before the stop or expiry. The selected single-option variant was evaluated on the July 2026 test block: 100,908 base candidates and 908,172 policy rows (nine policies per candidate).
| All-policy July test metric | FTO-1.5 |
|---|---|
| ROC-AUC | 0.8808 |
| Average precision | 0.5951 |
| Brier score | 0.0822 |
| Log loss | 0.2717 |
| Calibration ECE, 15 bins | 0.0221 |
For the shared fixed +1.0R / −0.5R policy, the direct July comparison is:
| Method | Test ROC-AUC | Top 1% hit / mean R | Top 5% hit / mean R | Top 10% hit / mean R |
|---|---|---|---|---|
| FTO-1 | 0.8735 | 88.4% / +0.826R | 64.5% / +0.468R | 51.6% / +0.273R |
| FTO-1.5 | 0.8553 | 86.7% / +0.801R | 69.4% / +0.541R | 52.2% / +0.282R |
On 2026-08-11, the frozen FTO-1 artifact was run against every graded record with an entry date on or after 2026-08-06. The run wrote 13,263 versioned prediction records to the monitoring store, and is idempotent, so re-running it cannot inflate the cohort.
| Forward-cohort status | Count |
|---|---|
| Scored records | 13,263 |
| Resolved TP/SL outcomes | 4,834 |
| Eligible resolved outcomes | 4,772 |
| Still in flight | 8,403 |
| Ungradeable | 26 |
| Entry dates with resolved outcomes | 2026-08-06 to 2026-08-07 |
The model was locked for this release-period cohort before the 2026-08-06 entries. Replaying that frozen artifact against the immutable entry-time snapshots gives the following candidate-level result on the 4,772 resolved, policy-eligible candidates:
| Forward reconstruction metric (raw score) | Result |
|---|---|
| Target-hit base rate | 30.32% |
| ROC-AUC | 0.9971 |
| Average precision | 0.9931 |
| Brier score | 0.0489 |
| Log loss | 0.1687 |
| Calibration ECE, 15 bins | 0.1019 |
| Mean predicted probability | 23.41% |
| Top 5% target-hit rate / mean R | 100.0% / +1.000R |
| Top 20% target-hit rate / mean R | 100.0% / +1.000R |
No FTO-1.5 forward claim exists yet. Its selected artifact was frozen on 2026-08-10, remains in shadow evaluation, and does not yet write versioned live predictions to the monitoring store. Refreshing the grading pipeline by itself cannot create prospective FTO-1.5 evidence.
A forward figure for FTO-1.5 will appear here only once its predictions are captured at recommendation time and graded against outcomes that occur afterwards, over a test month it has never been evaluated on, with costs and overlap included.
