You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Kalshi Perps Direction Classifier

Purpose: predict whether a crypto perpetual future's ("perp") price on Kalshi will be higher or lower 1 minute from now. A binary classifier trained on papylove/kalshi-perps-data.

Part of an open-source, dry-run-by-default automated trading bot: papykabukanyi/bettor. The bot never relies on this model alone -- an entry always also requires an independent technical scalper signal (a local price dip/rally against a trend filter) to agree first.

The prediction horizon changed from 30 minutes to 1 minute. The strategy checks every open position every 20 seconds and exits within minutes via take-profit/quick-profit/stop-loss, so a model answering "up or down 30 minutes from now" was answering a materially different question than what the strategy actually needs moment to moment.

Model details

Type: scikit-learn classifier, selected by comparing Logistic Regression, Random Forest, and Gradient Boosting on a chronological holdout split (never randomly shuffled -- this is time-series data, and a random split would leak future information into training).
Retrained: daily, on the full accumulated history in the companion dataset -- both the live Kalshi archive and (when available) the deeper Coinbase-sourced pretraining rows, see the dataset card.

Features

Every timeframe from 1 minute up is derived directly from the 1-minute close series -- Kalshi's and Coinbase's candlestick APIs both cap out at 1-minute native resolution, so nothing finer (5s/15s/30s) exists to train on from any free source.

ret_1m, ret_3m, ret_5m, ret_10m, ret_15m, ret_30m,   # 1-30 minute returns
trend_1h, trend_2h, trend_3h, trend_4h,               # 1-4 hour returns
dist_to_ma_15, dist_to_ma_30,   # distance from the 15/30-min moving average
volatility_5, volatility_15, volatility_30,   # rolling volatility, 3 scales
trend_pct,                      # 6-hourly-candle trend
sentiment_score,                # news sentiment for the underlying coin
ticker_code                     # which instrument (categorical, see meta)

ticker_code is a label-encoding over the instruments present at training time -- see perps_model_meta.json's ticker_categories list for the exact mapping (order matters: index = code).

Performance (most recent training run)

Chronological holdout, ~150k rows (capped for memory -- see PERPS_MAX_TRAIN_ROWS in the GitHub repo), 13 instruments, 1-minute label horizon:

Model	Accuracy	ROC AUC
Logistic Regression	0.517	0.529
Random Forest	0.536	0.548
Gradient Boosting (selected)	0.537	0.549

Read this honestly: an AUC of ~0.549 is a real but modest edge over random (0.5) -- not a strong predictive signal on its own, comparable to what the previous 30-minute-horizon version of this model showed. This is exactly why the bot requires the technical filter to agree too, and why a separate backtest (perps_backtest.py in the GitHub repo) validates the combined strategy's trade frequency and win rate, not this model's accuracy in isolation. See perps_model_meta.json's scores field on this repo for the current numbers, updated on every retrain.

Usage

import joblib
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("papylove/kalshi-perps-model", "perps_model.joblib")
model = joblib.load(model_path)

# x is one row, in the exact order of perps_model_meta.json's feature_columns
proba_up = model.predict_proba([x])[0][1]

Limitations

No Kalshi trading fees or bid-ask slippage are modeled anywhere in this project's backtests -- treat any backtested return figures as optimistic.
Kalshi's crypto perps are a new product (first listings 2026-06-04), so the LIVE (traded) training data covers at most ~7 weeks per instrument. The dataset's Coinbase-sourced rows extend this for pretraining, but that is the underlying coin's spot price, not Kalshi's own perp quote -- a reasonable proxy given a perp tracks its spot price closely, not a guarantee of identical behavior.
Confirmed live: many of Kalshi's immediate_or_cancel orders don't fully fill at the requested price -- this model (and any backtest built on top of it) has no way to account for real order-book liquidity/fill-rate, so actual trade frequency and returns should be expected to run below any backtested figure.
Not financial advice. Provided for research/educational purposes as part of an open-source project.

Downloads last month: -; Downloads are not tracked for this model. How to track