New Year 2027 Special$9.90/mo$19Save 48%Get New Year Offer →
Official benchmark

Model Accuracy — Official Benchmark

HPSILab provides market analysis, not trading decisions. This benchmark measures how the model behaves under a fixed test protocol. It is not a promise of future performance and should not be interpreted as a buy or sell signal.

The official accuracy comes from one fixed benchmark: the same 18 stocks, predicted once per NYSE session at 16:15 New York time and scored at the next session's official close. Every call stays on the record — right, wrong or pending.

16:15 ET, every NYSE sessionNo deletions, no rewrites

What is — and is not — counted

Only Benchmark predictions are scored here. Predictions you request any time on the website, the REST API, the Python SDK or MCP are separate and never enter this figure, no matter how many there are.

Accuracy
…
awaiting first scored session
Brier score
…
lower is better · 0.25 = coin flip
Evaluated
…
Correct: 0
Pending
…
Total predictions: 0
"Always up" baseline
…
same sessions, guessing up every time
Evaluation period
…
Latest run
…
void
…
0 void (corporate action, not scored)

Per-stock accuracy

TickerTotal predictionsEvaluatedPendingCorrectAccuracyBrier score
…

Calibration

When the model says X% chance of an up close, how often does it actually close up?

P(up) rangenAvg predictedActual up rate

How the Brier score is calculated

BS=1NN∑i=1(pi − oi)2

N = number of scored predictions · pᵢ = the model's P(up) for prediction i · oᵢ = 1 if the target close is higher than the reference close, otherwise 0 (flat counts as 0).

Accuracy only checks the direction. The Brier score also checks the confidence: a confident call that turns out right scores close to 0, a confident call that turns out wrong is penalised heavily. Range 0 (perfect) to 1 (always fully confident and wrong). Always answering 50% scores exactly 0.25, so anything below 0.25 is better than a coin flip.

Example

  • P(up) = 70%, closed up: (0.70 − 1)² = 0.09
  • P(up) = 70%, did not close up: (0.70 − 0)² = 0.49
  • P(up) = 50%, either outcome: (0.50 − 1)² = (0.50 − 0)² = 0.25

The headline figure is the average over every scored prediction; per-stock figures average that stock's predictions only.

Recent benchmark predictions(every call, including misses and pending)

Download CSV (all)
Predicted onTarget sessionTickerCallP(up)Reference closeTarget closeResultGenerated (ET)

How the benchmark works

  • Universe: the AI Stocks and Semiconductor Stocks lists from our US markets page — 18 unique tickers (NVDA and AMD are in both and count once).
  • Timing: a scheduled job runs at 16:15 America/New_York on every NYSE session, after the close, using that day's complete data. Predictions are accepted only inside 16:15–16:45; a missed window is never filled in later.
  • Target: the next real NYSE session from the exchange calendar — Friday predicts Monday, and a holiday Monday moves the target to Tuesday.
  • Scoring: after the target session's official close, the close is compared with the reference close. Up = higher; an unchanged close counts as not up. Brier score = (P(up) − outcome)².
  • Audit: a prediction cannot be edited or deleted once written — the database refuses it. A stock split between the two closes makes the pair incomparable; that row is kept and marked void.
  • The benchmark began clean on launch day. No earlier predictions were copied into it.

Past accuracy does not guarantee future results. This page is for informational purposes only and is not investment advice.

Market conditions can change because of new information, policy changes, earnings, geopolitical events, and other factors that may not be known at the time of analysis.

HPSILab does not provide investment advice or make investment decisions for users.

Model Accuracy — Official Benchmark | HPSILab