HPSILab Market Direction
A probabilistic, forward-tracked market-direction decision model for AI trading agents.
Overview
HPSILab Market Direction estimates the probability that a US equity will close higher on the next US trading session. It answers one typed question with a probability, so an agent can use the number directly — size a position by it, gate a trade on it, or combine it with its own signals — instead of parsing prose.
What sets it apart is the record behind the probabilities. HPSILab runs a public forward benchmark: predictions are made before the outcome exists, stored unchanged, scored at the next official close, and published with Brier score and a calibration table, wrong calls included.
Publisher
HPSILab
HPSILab builds market decision infrastructure for AI trading agents.
Website: https://hpsilab.com/
Decision primitive
noul
Will {ticker} close higher on the next US trading session?
Input
| Field | Where | Notes |
|---|---|---|
ticker | URL path | US equity or ETF symbol, e.g. NVDA. A malformed symbol returns 400 with a suggested correction in the X-HPSILAB-Suggested-Symbol header; a well-formed symbol with no data returns 404. |
There are no other inputs. The model reads the latest daily market state for the ticker itself.
curl https://hpsilab.com/api/ai_prediction/NVDA \
-H "Authorization: Bearer $HPSI_API_KEY"
The same prediction is available to agents as the get_ai_prediction tool on the HPSILab MCP server (https://hpsilab.com/mcp). Authentication is an hpsi_ API key sent as a Bearer token, or OAuth 2.1 for connectors that support it.
Output
The REST endpoint returns a JSON array with one object. The MCP tool returns the same object without the array.
| Field | Type | Meaning |
|---|---|---|
ensemble_up_probability | number, 0–1 | The decision. Probability that the next session closes higher. |
rf_up_probability | number, 0–1 | Random-forest component of the ensemble. |
lr_up_probability | number, 0–1 | Logistic-regression component of the ensemble. |
last_date | string | Date of the last daily bar the prediction was made from. The prediction is for the session after it. |
last_close | number | Close on last_date. |
model_variant | string | Identifier of the model variant that produced it (currently candidate). |
daily_vol_est | number or null | Estimated daily volatility. |
suggested_stop_loss_price | number or null | Volatility-based reference level. |
suggested_take_profit_price | number or null | Volatility-based reference level. |
sentiment_score | number or null | News sentiment input, when available. |
Example
Illustrative only — not a real prediction.
[
{
"last_date": "2026-10-02",
"model_variant": "candidate",
"last_close": 100.0,
"ensemble_up_probability": 0.54,
"rf_up_probability": 0.56,
"lr_up_probability": 0.52,
"daily_vol_est": 0.021,
"suggested_stop_loss_price": 97.9,
"suggested_take_profit_price": 104.2,
"sentiment_score": 0.1
}
]
Evaluation
What is evaluated
HPSILab publishes an official forward benchmark at hpsilab.com/accuracy:
- Benchmark: US AI + Semiconductor Benchmark (
us-ai-semiconductor-v1) - Universe: 18 fixed tickers — NVDA, MSFT, GOOGL, META, AMZN, PLTR, AMD, ORCL, AI, SOUN, AVGO, QCOM, INTC, MU, TXN, MRVL, AMAT, ASML
- Schedule: one prediction per ticker at 16:15 New York time on every NYSE session, for the next NYSE session
- Scoring: at the next official close — direction, correct/incorrect, Brier score
- Record: predictions are written once and never edited; only predictions made inside the 16:15–16:45 window count
The benchmark scores a separate, dedicated benchmark model (benchmark-logit-v1, a pooled logistic regression), not the production model behind this API. It is kept separate so that the scoreboard cannot be tuned and so that every stored probability can be re-derived from the same price bars. Read its numbers as the forward record of HPSILab's direction modelling, not as a measurement of the exact responses this endpoint returns.
Results
The benchmark started on 2026-10-03. Headline numbers are published once at least 100 predictions have been scored — the same rule the accuracy page applies — and will be added here at that point:
| Metric | Value |
|---|---|
| Accuracy | pending (fewer than 100 scored predictions) |
| Always-up baseline accuracy | pending |
| Brier score | pending |
| Calibration error (ECE, over the benchmark's calibration buckets) | pending |
| Calibration table (predicted vs. actual up-rate per bucket) | pending — live on hpsilab.com/accuracy |
| Scored predictions | 0 as of 2026-10-04 |
Latency
End-to-end HTTP latency of successful GET /api/ai_prediction/{ticker} requests in production over the 30 days to 2026-10-04 (2,655 requests, none excluded):
| Median | p95 |
|---|---|
| 254 ms | 1,534 ms |
This is server-side request time, measured by HPSILab's request logging. A ticker's first call for a new trading day computes the prediction; later calls return the stored one and are faster. Both are included.
Why accuracy alone is insufficient
Next-session stock direction is close to a coin flip, and the share of up days varies by stock and by period. 80–90% accuracy figures from ordinary classification tasks are not a meaningful comparison. Judge this model on:
- Accuracy against the always-up baseline over the same scored sessions — beating 50% means little if the stocks rose 55% of the time.
- Brier score — rewards probabilities that are both right and appropriately confident.
- Calibration — when the model says 60%, does the stock go up about 60% of the time?
- An out-of-sample forward record — predictions fixed before the outcome, not a backtest.
Intended use
- AI trading agents that need a probability rather than a narrative
- Research and signal evaluation
- Probabilistic decision systems that combine several inputs
Limitations
- The output is a probability, not a guarantee; a high-probability call can still be wrong.
- Market regimes change, and a model fitted on past behaviour can degrade when they do.
- Direction accuracy ignores the size of moves and transaction costs; a profitable strategy needs more than a correct direction.
- The published benchmark scores a dedicated benchmark model on 18 AI and semiconductor stocks, not the production model across every ticker.
- For research and informational use. Not investment advice.
Availability
Hosted API only. Weights are not published.
Links
- Main site: hpsilab.com
- Developer docs: hpsilab.com/developer/v2
- MCP server:
https://hpsilab.com/mcp - Accuracy / benchmark: hpsilab.com/accuracy
- Pricing: hpsilab.com/pricing