BehaviorBench / docs /schema_trade.md
mingzhu0527's picture
Release BehaviorBench artifact (from lyang84/BehaviorBench-camera-ready)
f9b150b verified
|
Raw History Blame Contribute Delete
2.37 kB

Trade data and label semantics

What a row represents

Trade data models a wallet's next observed transaction. Structured split rows include wallet, market, block/log position, transaction side, position before and after the transaction, derived action labels, and market/price features. Model-ready benchmark rows convert this context into text and a structured target.

Principal source files

  • dataset/Trade/source/trade_price_matched.parquet: upstream price-matched trade table.
  • dataset/Trade/source/behavior_fixcost.parquet: 11,531,670 rows with behavior, position, PnL, and K-line features.
  • dataset/Trade/source/market_semantic.parquet: market semantic metadata.

Important structured columns

Group Columns Interpretation
Identity wallet, market_id, outcome Wallet and traded outcome.
Market text event_slug, question, market_slug Semantic context.
Ordering block, log_index, tx_hash On-chain event location and identifier.
Trade side, amount_int, delta_pos Observed trade side, amount, and position delta.
Action labels action_label, behavior_label_fine Derived behavioral targets.
Position prev_pos, new_pos, position_size, position_change Position state and change.
Sequence seq_idx_outcome, seq_idx_market, gap and previous-trade fields Within-wallet behavioral ordering features.
Open/close/flip is_open_trade, is_close_trade, flip_*, switched_* Derived state-transition indicators.
Value/PnL cost_basis_avg, mark_price, unrealized_pnl_* Position valuation features.
K-line k0_* through k5_* Released multi-window market-price features.

The structured split files additionally include prior_same_market_trades.

Model-ready target

Direct benchmark files contain:

sample_id, task_type, split, wallet, input_text, target, output_format

Representative targets request:

{"direction": "increase | decrease | close", "amount_usdc_est": 0.0}

Interpretation:

  • increase: add exposure to the current outcome;
  • decrease: reduce exposure without the released prompt treating it as a full close;
  • close: close the position.

The released benchmark uses these derived action and amount labels consistently across its model-ready interfaces.