OpenBabylon/unlp-ukraine-forecasting

UNLP Ukraine Forecasting Dataset

A prediction-market forecasting benchmark with 111 resolved Ukraine-related Polymarket markets, 672 frozen cutoff points, 7 fixed prediction horizons (6h, 12h, 1d, 2d, 3d, 5d, 7d), and ~660K model predictions from 5 LLMs (GPT-5, GPT-5-mini, Gemini 2.5 Flash, Gemini 2.5 Pro, Gemini 3.1 Pro Preview) under 4 information conditions (A/C/D/D_UA).

Built by walking the same code path as the UNLP 2026 evaluation harness (eval/harness.py:build_prompt()) — every prompt and ground-truth label is byte-identical to what was used to elicit the model predictions.

What's in this repo

data/                       The benchmark itself (prompts + ground truth + baselines)
  unlp-ukraine-A-train.jsonl       — 529 examples (89 markets × ~6 cutoffs)
  unlp-ukraine-A-test.jsonl        — 143 examples (22 markets × ~6 cutoffs)
  unlp-ukraine-C-train.jsonl, ...
  unlp-ukraine-D-train.jsonl, ...
  unlp-ukraine-D_UA-train.jsonl, ...
  Total: 2,688 examples (672 cutoffs × 4 conditions)

predictions/                Model predictions with full reasoning chains
  predictions_ukraine_v4_gemini-2.5-pro.csv         — 106,079 predictions
  predictions_ukraine_v4_gemini-2.5-flash.csv       — 156,754
  predictions_ukraine_v4_gemini-3.1-pro-preview.csv —  70,137
  predictions_ukraine_v4_gpt-5.csv                  —  84,659
  predictions_ukraine_v4_gpt-5-mini.csv             —  84,338
  predictions_ukraine_v5_gemini-2.5-pro.csv         —  43,556
  predictions_ukraine_v5_gemini-2.5-flash.csv       —  71,792
  predictions_ukraine_v5_gpt-5-mini.csv             —  45,210
  Total: ~660K predictions

metadata.json
README.md

Conditions

Each cutoff is rendered under 4 information conditions:

Condition	Context provided
A	Price summary only (blind baseline)
C	A + news context (48h/7d/30d article windows)
D	A + chart + news + cross-market context + comments
D_UA	D + Ukrainian sources (Telegram bloggers, casualties, Militarnyi)

Horizons

Fixed 7 prediction horizons from each cutoff:

Label	Hours forward
6h	6
12h	12
1d	24
2d	48
3d	72
5d	120
7d	168

Ground truth is the actual market price at each horizon timestamp (6-hour tolerance for nearest available price point).

Benchmark schema (data/)

Each row in unlp-ukraine-{A,C,D,D_UA}-{train,test}.jsonl:

Identity: market_id, question, event_slug, cutoff_timestamp, hours_remaining, condition, split

Prompt (model input): prompt (full rendered V4 prompt, byte-identical to what was sent to the LLMs), has_chart, has_war_map

Ground truth: current_price, labels (dict mapping horizon labels to actual prices: {"6h": 0.35, ..., "7d": 0.42}), outcome (binary), end_date

Baselines: baseline_nochange (predict current price unchanged), baseline_momentum (7-day linear regression extrapolated)

Volume: volume_at_cutoff_24h (USD notional in 24h before cutoff)

Predictions schema (predictions/)

Each row in a predictions_*.csv file:

Column	Description
market_id, event_slug, question, category	Identity
cutoff_timestamp, condition, horizon	(Join key to benchmark — see below)
regime	Market lifecycle bucket: early, mid, late, final_2w
current_price	Market price at cutoff (model input)
predicted_price	Model's predicted price for this horizon
actual_price	Ground truth price at cutoff + horizon hours
error	`
baseline_nochange_error	MAE of no-change baseline at this row
baseline_momentum_error	MAE of momentum baseline at this row
direction_correct	Bool: did the model predict the correct direction?
hours_to_resolution	Hours from cutoff to market resolution
reasoning	Full chain-of-thought reasoning from the model
background_assessment	Model's background analysis (V4 prompt structure)
news_interpretation	Model's interpretation of provided news
prediction_reasoning	Model's stated rationale for its number
confidence	high / medium / low
model	Model name (e.g., gemini-2.5-pro)
prompt_version	v4 (absolute price) or v5 (Bowman framework)
context_quality, has_48h_news, unexplained_move	Diagnostic flags

Joining predictions with the benchmark

Each prediction row has (market_id, cutoff_timestamp, condition, horizon) which uniquely identifies the corresponding benchmark row (the benchmark's labels dict is the source of actual_price):

import pandas as pd, json

# Load benchmark for one condition
benchmark = [
    json.loads(l)
    for l in open("data/unlp-ukraine-D-train.jsonl")
]

# Load predictions for one model
preds = pd.read_csv("predictions/predictions_ukraine_v4_gemini-2.5-pro.csv")
preds_D = preds[preds.condition == "D"]

# Join key: (market_id, cutoff_timestamp, horizon)
# benchmark row → 7 prediction rows (one per horizon)

Stats

	Count
Markets	111
Frozen cutoffs	672
Train / test markets	89 / 22 (temporal split by resolution date)
Benchmark examples (cutoffs × conditions)	2,688
Model predictions	~660K
Models evaluated	5 (GPT-5, GPT-5-mini, Gemini 2.5 Flash/Pro, Gemini 3.1 Pro Preview)
Prompt versions	v4 (absolute price), v5 (Bowman aspiration framework)
Markets with 24h volume > 0	99%

Scoring

MAE (Mean Absolute Error): |predicted_price - actual_price| at each horizon. Compare against:

• No-change baseline: predicted = current_price
• Momentum baseline: 7-day linear regression extrapolated forward

direction_correct provides a secondary directional-accuracy metric.

Source

From the OpenBabylon/signal repo. Built using eval/build_unlp_dataset.py (benchmark) and eval/harness.py (predictions). The harness uses the V4 prompt template at eval/prompts/v4/system.txt and task.txt.

License

MIT.