PLTR forecast model
and what it got wrong.
Palantir Technologies Inc. is one of 25 stocks GioModel models. This page is the public record: the inputs, the method, the evidence gates, the validation metrics, and past forecasts graded against what actually happened. Current frozen ranges are the only part held for members.
Company overview
- Company
- PLTR Palantir Technologies Inc.
- Sector
- Information Technology vs XLK
- Market cap
- 380.7B
- Model status
- 3 / 9 evidence gates passed
Reference fundamentals as of Aug 5, 2026: gross margin 84.8%, operating margin 42.8%, P/E 139.12, P/S 61.85. These are context for reading the model, not model inputs — the admitted inputs are listed below. Price history comes from Nasdaq.
How the PLTR forecast is generated
An interval model: it estimates a distribution of prices at 5, 10 and 21 sessions ahead from a frozen history, then reports the 10th, 50th and 90th percentiles. Selection, calibration and terminal evaluation happen in that order with a strict 21-session embargo between them, so the block used to score the model was never used to choose it.
- 01Freeze the history
A point-in-time panel of completed PLTR sessions is fixed and hashed — this release froze on Jul 31, 2026. Nothing after that date can influence the numbers, which is what makes a later grade meaningful.
- 02Build inputs that could have existed
Every feature is computed from a strictly prior window and tested for prefix invariance, so a value never changes when later data arrives. The admitted inputs are listed below.
- 03Fit and score through time
Position-range model v1 candidates are compared inside purged, embargoed walk-forward splits across 252 evaluation origins from Jul 1, 2025 to Jul 1, 2026. The block used to score a model is never the block used to pick it.
- 04Publish, then grade in public
The release states which of the nine gates it passed and which it did not. Thirty days later the frozen numbers appear on the public track record and are graded on the exact trading session the horizon lands on.
What data actually goes in
Not everything a model could use is admitted. Each input is marked with how far it has actually been taken — admitted means it feeds the published forecast; anything else is disclosed rather than quietly implied.
| Category | Input | Status |
|---|---|---|
| Fundamental | Revenue growth | PLANNED |
| Fundamental | Margins | PLANNED |
| Fundamental | Earnings | PLANNED |
| Fundamental | Valuation | PLANNED |
| Technical | Momentum | CONSTRUCTED |
| Technical | Moving averages | CONSTRUCTED |
| Technical | Volatility and volume | CONSTRUCTED |
| Market | Interest rates | PLANNED |
| Market | Sector trends | CONSTRUCTED |
| Market | Market regime | CONSTRUCTED |
| Portfolio | Correlation | PORTFOLIO LAYER |
| Portfolio | Position size | LIVE ARITHMETIC |
| Portfolio | Risk contribution | PORTFOLIO LAYER |
Validation metrics
Historical terminal development comparison; not post-protocol forward evidence.
- Evaluation origins
- 252
- Direction accuracy
- 47.5% vs 52.5% always-one-class
- 80% band coverage
- 95.0%
- Mean WIS vs baseline
- −0.0% lower is better
A historical win means the forecast direction matched the realized direction at the stated horizon. It is not a trading return. Horizons scored: 5, 10, 21 sessions. Fold record: 0 of 4 on mean weighted interval score versus expanding baseline. Calibration error 7.64%. A win here is a scoring outcome on historical data, never a trade.
Evidence gates: 2 of 9
3 gates in the frozen release cannot be graded, so they are no longer counted as demonstrated. The frozen artifact itself is unchanged — this corrects what we claim it proves.
- G3 Interval score — The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
- G4 Median error — The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
- G6 Direction calibration — The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
The frozen release passes 3 of 9 locked gates. One invalid source row is quarantined and the expanding baseline remains the fail-safe controller.
Every gate below is reproduced from PLTR's own frozen release artifact — the rule as written before the results were seen, the measured value, and the verdict. Gate numbers are local to this protocol: each model family predeclared its own set, so a G3 here is not a G3 on another model's page. Compare the protocols side by side.
This model is graded out of 9, not out of a single site-wide number. This family still publishes a directional claim, so it keeps the gate that tests it and is graded out of nine. The position-forecast family dropped that gate on 15 August 2026 together with the claim itself, and is graded out of eight — a gate is only ever removed alongside the claim it was there to check.
- G1Data quality — open
Frozen threshold: >=504 sessions; zero duplicate dates and invalid target OHLCV rows
Observed: 1465 sessions; 1 invalid target OHLCV rows
- G2No lookahead — passed
Frozen threshold: prefix invariance PASS and zero incomplete training outcomes
Observed: prefix invariance PASS; zero incomplete training outcomes
- G3Interval score — not gradeable
Not gradeable: The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
Frozen threshold: selected WIS <= 98% of calibrated expanding baseline
Observed: selected 0.452250; baseline 0.452250
- G4Median error — not gradeable
Not gradeable: The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
Frozen threshold: selected median absolute error <= 102% of calibrated baseline
Observed: selected 0.067235; baseline 0.067235
- G5Coverage — open
Frozen threshold: aggregate 0.72-0.88; every horizon 0.68-0.92
Observed: aggregate 0.950; horizons 0.869, 0.992, 0.988
- G6Direction calibration — not gradeable
Not gradeable: The published candidate is the baseline itself, so this gate compares the model against a copy of itself. No comparison was demonstrated either way.
Frozen threshold: selected Brier < expanding baseline Brier
Observed: selected 0.255661; baseline 0.255661; ECE 0.0764; probability std 0.0083
- G7Fold consistency — open
Frozen threshold: >=3/4 chronological folds
Observed: 0/4 chronological folds beat baseline
- G8Reproducibility — passed
Frozen threshold: same-seed exact; maximum displayed quantile range <= 0.005
Observed: same-seed exact True; max seed range 0.000795
- G9Independent forward evidence — open
Frozen threshold: >=30 non-overlapping post-protocol outcomes spanning >=90 calendar days
Observed: 0 eligible new outcomes across 0 calendar days
Historical Evaluation Ledger — Development/OOS Evidence
These origins were evaluated during model development, before the public protocol began. They are out-of-sample within the development panel, but they were not frozen and published ahead of their outcome, so they cannot demonstrate what this model does on genuinely unseen data. Treat them as construction evidence.
The forward record — releases frozen and hashed before the market moved, then graded on their exact resolution session — is kept separately on the Public Forward Track Record. Nothing there has been disclosed yet: the first unlock is 2026-07-29.
Past evaluation origins for PLTR, each graded against the realized move. Only origins older than 0 days appear; anything newer sits behind the same delay as the live release.
| Origin | Horizon | Forecast median | Realized | Direction | Result |
|---|---|---|---|---|---|
| 2026-07-01 | 21 sessions | +2.26% | −2.12% | UPactual DOWN | LOSS |
| 2026-07-01 | 10 sessions | +0.82% | +6.93% | UPactual UP | WIN |
| 2026-07-01 | 5 sessions | +0.70% | +2.63% | UPactual UP | WIN |
| 2026-06-30 | 21 sessions | +2.32% | +4.79% | UPactual UP | WIN |
| 2026-06-30 | 10 sessions | +0.85% | +14.65% | UPactual UP | WIN |
| 2026-06-30 | 5 sessions | +0.68% | +13.33% | UPactual UP | WIN |
| 2026-06-29 | 21 sessions | +2.32% | +6.31% | UPactual UP | WIN |
| 2026-06-29 | 10 sessions | +0.87% | +15.57% | UPactual UP | WIN |
| 2026-06-29 | 5 sessions | +0.70% | +16.14% | UPactual UP | WIN |
| 2026-06-26 | 21 sessions | +2.33% | +9.39% | UPactual UP | WIN |
| 2026-06-26 | 10 sessions | +0.88% | +15.15% | UPactual UP | WIN |
| 2026-06-26 | 5 sessions | +0.72% | +17.36% | UPactual UP | WIN |
The Jul 31, 2026 PLTR release is now public in full on the track record, graded on its exact resolution sessions. That delay is the whole product boundary: a forecast that is already thirty days old cannot be traded on profitably, but it can be checked — which is what makes the record worth reading. See every release and its grade.
NO-GO does not mean the research contains no information. It means GioModel refuses to convert incomplete evidence into a trading recommendation. The PLTR model publishes its ranges, probabilities and diagnostics to members, and publishes every gate it has not passed to everyone — but it does not claim to have earned directional deployment, and the deployment gate is independent of payment. Nothing on this page is investment advice, a price target, or a recommendation to buy or sell Palantir Technologies Inc.. Read the full method and the research disclaimer.