PIONEER / PROBABILITY MODELS

Pricing
the unpriced.

Custom probability models for difficult-to-price events. Pioneer is being developed to turn evidence into estimates and two-sided quotes for the questions conventional models leave open.

In development at Obsidian Cube

The markets nobody else prices

01 · WHAT PIONEER IS

A pricing desk for the markets
that never get a price.

Prediction markets turn questions into contracts that pay 100¢ for yes and 0¢ for no. The famous ones trade deeply; thousands of others never get a serious price. Pioneer builds a model for each of those, posts the first bid and ask, and keeps the price honest as evidence arrives.

02 · TRAINING LAB

A signal takes shape.

Choose a word. Train a small neural network on historical speech counts. Watch its connections change as it learns to estimate whether that word appears in the next minute.

EXPERIMENT 01 / LANGUAGE → PROBABILITY3 INPUTS / 6 HIDDEN NODES / 1 OUTPUT
Ready to learn
A small neural network learning from historical word countsThree input nodes connect to six hidden nodes and one probability output. Connection widths represent weight magnitude. Solid connections are positive; dashed connections are negative. OBSERVATIONSHIDDEN STATEESTIMATE
FIG. 08 / THE LEARNING FIELDA representative historical training window.Line weight = connection strength · node fill = activation

VALIDATION / NEXT MINUTE

Will “tariff” appear?

—%

A fixed historical example, held out from training.

Epoch
0 / 180
Training loss
—
Validation loss
—
Latest weight change¹
—

Ready. Training starts when you choose.

LOSS / LOWER IS A CLOSER FIT

TrainingValidation
Training and validation cross-entropy loss by epochThe curves are computed as this network trains. Validation data is never used to update its weights.
Inside the experiment

Actual learning. A small example.

A 3 → 6 → 1 neural network runs in your browser. Its three inputs are the previous three minute counts, transformed with log(1 + count). Six tanh units feed a sigmoid output: the probability of at least one occurrence in the following minute. Full-batch gradient descent updates all 31 parameters using cross-entropy loss, with a fixed step size of 0.12.

Historical windows are separated into training and validation sets.

This is an educational experiment on one speech, not Pioneer’s production model or a trading signal. A lower training loss does not establish that a model will generalize. Validation loss can rise. ¹ The displayed change is the root-mean-square update across the 24 drawn connection weights; the seven biases are also trained.

Data and visual key

Minute counts are derived from the official White House captions of the April 2, 2025 “Make America Wealthy Again” event. Captions may contain errors. The three minute bins touching the previously excluded excerpt and the final incomplete minute are omitted. No audio, video or streaming feed is used.

Wider lines carry larger absolute weights. Solid lines are positive; dashed lines are negative. Node fill shows activation. During training, dots trace the forward pass and the backward update; their speed is illustrative. Connection curves change with the actual learned weights.

Initialization and sample order are fixed so a reset reproduces the experiment. The right-hand estimate always uses the same first validation window for the selected word; its outcome never enters training. The network diagram cycles through representative training windows, so its output can differ from the validation estimate.

Inspect the learned weights
Network parameters at the current epoch
ParameterValueLatest update

Educational experiment on one speech; not Pioneer’s production model or a trading signal. / Pioneer in development

03 · THE LEARNING PROCESS

Observe. Compare. Adjust.

A probability begins as an estimate. Learning changes the connections that produced it.

  1. 01 / FORWARD

    Follow the signal.

    Three historical counts pass through the network. Each connection weights the signal before it becomes a probability.

  2. 02 / COMPARE

    Measure the error.

    The estimate meets the known outcome. Cross-entropy loss measures how much probability the model assigned to what happened.

  3. 03 / BACKWARD

    Change the weights.

    The error travels backward. Gradients adjust each connection; untouched validation windows test whether the fit carries beyond training.

04 · THREE REAL MARKETS

One politics market, two culture markets.
Priced from the evidence.

Historical Kalshi market snapshots captured on October 22, 2026. This separate illustrative model scores the evidence, turns it into a quote, and compares it with the recorded exchange price. The snapshots do not update.

ILLUSTRATIVE FEATURE MODEL · REAL MARKETSQuestions and prices from Kalshi, Oct 22, 2026 · evidence weights illustrative · not a trading recommendation

Evidence behind the price

Contribution to the model's log-odds

    POLITICS MARKET · KALSHI

    Will the SAVE Act (H.R. 22) become law before Jan 4, 2027?

    —%

    Pioneer model probability of a yes resolution

    PIONEER BID—¢
    SPREAD—¢
    PIONEER ASK—¢
    KALSHI ON OCT 22, 2026 · ——¢ bid · —¢ ask—
    View on Kalshi

    Log-odds — · a market-specific base rate of — plus each bar above

    Market questions, resolution dates and exchange prices come from Kalshi's public market data as of October 22, 2026 and will have moved since. Evidence weights are illustrative; a production model learns them from resolved markets. Nothing here is a recommendation to trade. Obsidian Cube is not affiliated with Kalshi.

    05 · LIQUIDITY

    A model is only useful
    if someone will trade on it.

    Pioneer is being designed to act as the first liquidity in markets that would otherwise have none: resting quotes on both sides, sized to what the model knows and repriced as it learns.

    • Before: a 12¢ bid, a 78¢ ask, and nothing in between. Nobody can trade at a fair price.
    • After: Pioneer rests 41¢ and 47¢ in size. The market has a price, and a trader on either side can act on it.
    Illustrative order books. Pioneer's quotes tighten the market from 66¢ wide to 6¢ wide.

    06 · WHO IT IS FOR

    What we offer.

    Pioneer is being designed for the people who need a price in markets that have never had one.

    • EXCHANGESLaunch liquidity for new listings and the long-tail categories in-house desks skip, repriced continuously while the event unfolds.
    • MARKET CREATORSA fair-value estimate before listing, a resting market so early traders meet a price, and guidance on resolution rules a model can price.
    • TRADERSA real price instead of an empty book: tighter spreads, quotes that move with the evidence, and size on both sides.
    • PARTNERSPricing research on request. Bring a question no one quotes yet and get models, backtests and a plan to make the market.

    A few things to know.

    Is Pioneer available today?

    Pioneer is a new product in development at Obsidian Cube. This page shows its planned capabilities. The labs run real models on this page; they do not connect to exchanges or live order books.

    How does Pioneer manage risk?

    Spreads widen when the model is uncertain, quotes skew away from positions Pioneer has already accumulated, and exposure per market is capped. Models are only trusted with size after they have been scored on resolved markets they never saw in training.

    Is what the labs show real?

    The training lab runs an actual small neural network on historical word counts derived from official White House captions. It is an educational example on one speech, not Pioneer’s production model; captions may contain errors. The three markets in the evidence lab are real Kalshi markets with prices captured on October 22, 2026 and illustrative evidence weights; Obsidian Cube is not affiliated with Kalshi. Nothing on this page is a live feed, a real order, a trading recommendation or an endorsement.

    How is Pioneer different from Prism and Counterspan?

    Pioneer prices and makes markets in the long tail of hard-to-value questions. Prism aggregates many existing markets into clearer business intelligence, and Counterspan explores portfolios of event contracts shaped by business exposure. They are distinct products in development at Obsidian Cube.

    PIONEER / CONTINUE THE CONVERSATION

    Let’s compare notes.

    For research, data, and collaboration inquiries.