Methodology

How Prismora generates, evaluates, and tracks predictions for Polymarket traders

This page documents the data flow behind the dashboard, from market input to model output, resolution, evaluation, and simulated performance.

1 / Overview

Prismora is built for traders on prediction markets like Polymarket and Predict who want a model-based directional signal before each window opens. The workflow is fully transparent: every prediction is locked in before the window starts, every outcome is recorded after the window closes, and the full history is publicly visible.

The system receives recent Binance K-line data, passes it to an AI forecasting model, converts the output to a directional bias, resolves the window against the actual market result, and updates accuracy and simulation metrics. Nothing is adjusted retroactively.

2 / Model Architecture

The forecasting layer is designed around a Transformer-based time-series model. It accepts ordered historical OHLCV data and produces a point estimate for a future close price. Prismora exposes public evaluation artifacts such as predicted close, predicted return, direction, confidence score, confidence range, and realized market result.

3 / Input Data

The input data comes from Binance public spot market endpoints. Before each prediction window, the worker fetches recent K-line records for the relevant symbol and interval. Each record contains open, high, low, close, volume, and quote amount values. The default lookback is 400 candles.

4 / Direction Classification

The model output is a predicted closing price. Prismora computes predicted return as:

predicted_return = (predicted_close - window_open) / window_open

That return is classified with interval-specific thresholds:

5m   0.03%
15m  0.05%
1h   0.10%

If the predicted return is smaller than the threshold in either direction, the output is considered Neutral Bias and excluded from directional accuracy.

5 / Confidence Score

Each prediction includes a 0 to 100 confidence score. When the model provides a usable confidence value, Prismora normalizes it to that range. When fallback logic needs to infer confidence, the score is derived from the distance between predicted return and the interval threshold.

6 / Settlement And Evaluation

After a prediction window closes, Prismora fetches the corresponding Binance K-line and reads the actual open and close values. If close is greater than or equal to open, realized direction is Up. If close is lower than open, realized direction is Down. A matched output counts toward directional accuracy; a not matched output counts against it. Neutral output remains visible but does not affect directional accuracy.

7 / Simulation Rules

Initial bankroll: $100
Resolved directional window: $1 fixed change
Matched output: +$1
Not matched output: -$1
Neutral output: $0

This rule creates consistent scorekeeping across assets and intervals. It does not represent real prediction-market odds, fees, spreads, liquidity, settlement mechanics, or execution availability.

For reference, if you were applying this signal to a real Polymarket market with a typical 50/50 odds structure, a sustained accuracy above 55% would generally outperform random selection over a large sample. The dashboard's historical match rate and calibration charts let you assess whether any given asset and interval has reached that threshold.

8 / Limitations

The model is based on historical price patterns and cannot anticipate sudden external events, exchange outages, regulatory announcements, protocol incidents, or other discontinuities. Shorter windows generally have lower signal-to-noise than 1-hour windows. The simulation omits fees, slippage, market depth, timing constraints, and venue-specific rules. Past accuracy and simulated performance do not guarantee future behavior.

9 / Data Freshness

The public dashboard refreshes overview data periodically and asset detail pages reload when the selected interval, period, or prediction page changes. Each row records its UTC window, creation time, resolution status, and market result so visitors can distinguish fresh pending output from resolved historical data.