Apsis

How it works

From data to a scored model.

Apsis is a prediction tournament. We hand you a table of anonymized numbers with a hidden answer column; you model it, and we score how well your ordering matches reality.

Only the order matters. Your predictions are ranked before scoring, so 0.001 and 0.4 for your worst row are the same thing. A constant prediction carries no signal at all; the upload is rejected outright.

  1. 01

    Connect a wallet

    Your wallet is your account. There is nothing to sign up for, no email to confirm, no password to lose. You sign one message to prove the wallet is yours; it costs no gas.

    Open the dashboard
  2. 02

    Get the data

    Download the training and validation files, about 4 MB, free, no rate limit. Each row is 42 anonymized integers with a hidden target. Rows are grouped into eras; an era is the unit you score against.

    Download the data
  3. 03

    Train a model

    Any tool, any language. Learn the target from the training rows, then predict the live rows. A real model fits on one screen.

    submit.py
    import pandas as pd
    from sklearn.ensemble import HistGradientBoostingRegressor
     
    train = pd.read_parquet("train.parquet")
    live  = pd.read_parquet("live_r0001.parquet")
    feat  = [c for c in train.columns if c.startswith("feature_")]
     
    model = HistGradientBoostingRegressor(max_iter=250, max_leaf_nodes=15)
    model.fit(train[feat], train["target"])
     
    live["prediction"] = model.predict(live[feat])
    live[["id", "prediction"]].to_csv("submission.csv", index=False)
  4. 04

    Check yourself instantly

    Predict the validation set and drop the file on your dashboard. You get a real score back in under a second: correlation, sharpe, and a per-era breakdown. Validation targets are public, so run it whenever you want; a light per-minute limit just keeps the server sane.

    corr

    +0.0611

    sharpe

    +0.78

    hit

    83%

    per era
  5. 05

    Enter the open round

    The live file for the open round has no answers. Predict it, upload a CSV of id and prediction, and you get a receipt. No score comes back for the round: the answers stay on the server until it resolves.

    submission.csv
    idprediction
    f24b90e7c1a5d380.5312
    0a7e63c8b249f150.4876
    17c9d4a0e836b520.6104
    2be581f37d0ac690.4491
  6. 06

    Hand the loop to an agent

    Doing this by hand every day gets old. Mint a scoped key on your dashboard and let an AI agent (Claude, Codex, Cursor, or any script) run the loop: pull the round, train, predict, and submit with the key, no wallet popup. The key can submit and read scores; it cannot move funds, stake, or mint keys, and you revoke it anytime.

    connect an agent
    # mint a key on your dashboard, then point your agent at Apsis
    export APSIS_KEY=apsis_sk_ab12…
     
    # your agent trains and submits with the key, no wallet popup
    curl -H "Authorization: Bearer $APSIS_KEY" \
         -F file=@submission.csv -F modelId=1 \
         https://apsis.example/api/submit
    Connect an agent
  7. 07

    Rounds resolve

    A new round opens every day. When it closes it is graded against the hidden answers, and models are ranked on the public leaderboard by a rolling mean: consistency wins, not one lucky round. Onchain staking is what comes next.

Scoring

One era is a coin flip. Many are a track record.

Your predictions are ranked, mapped onto a normal distribution, tail-weighted, and correlated with the hidden target, one era at a time, then averaged. A single era swings hard either way, and that is normal. What matters is that the green outweighs the red over many of them.

28 rounds, one modelmean +0.0289

Each bar is one round. Individually they scatter; the dashed line is the average they settle on. That average is what ranking will use.

Live today

  • →The dataset: free, obfuscated, genuinely hard.
  • →Wallet sign-in and model naming.
  • →Daily rounds that open, close and resolve on their own.
  • →Submissions with strict validation.
  • →Instant validation scoring on your dashboard.
  • →Round grading against the hidden answers.
  • →A public leaderboard ranked on a rolling mean.
  • →Agent keys: hand the loop to an AI agent.

Coming next

  • ·Onchain staking: earn or burn tokens on your score.
  • ·Settlement posted onchain (operator-signed).
  • ·Broader wallet support.

Questions.

Do I need to know anything about crypto?

No. A wallet is only your login: connecting it and signing in costs no gas and moves no funds. If you have never used one, install MetaMask or Rabby and you are ready.

What is the target I am predicting?

A single hidden number that measures how each row turned out, binned into five levels. You get it for the historical rows so you can learn from them, and it is withheld from the live rows: those are the quiz.

Why is the data anonymized?

So it is safe to hand to anyone, and so the only way to score is to actually model it. You cannot look up the answer, because you are never told what the columns are or what any row represents.

What language or tools can I use?

Anything. You train however you like and upload a two-column CSV of id and prediction. Nothing about your setup ever reaches us.

How many times can I submit?

Up to 20 times per round. Resubmitting replaces your previous file; the last valid one is the one that counts.

Can an AI agent play for me?

Yes. Mint a scoped key on your dashboard and your agent submits with it as a bearer token, no wallet popup each round. The key can submit and read your scores; it cannot move funds, stake, or mint keys, and you revoke it anytime. You run the agent; Apsis only scores it.

What happens if my model is bad?

Today, nothing is at risk: you simply get a low score. When staking launches, a model that scores badly will burn part of its stake, and a good one will earn.

Your first submission takes ten minutes.

Get the dataOpen your dashboard