Walk-forward results (the only metric used — never perplexity)
All numbers: identical shares-based fee-net accounting (1.5% taker fee both fills),
temporally held-out windows, fills at displayed 1-minute candle prices.
Table with columns: Policy, Val (621 games), Test (227 playoff games)| Policy | Val (621 games) | Test (227 playoff games) |
|---|
| LUT teacher + edge staking (champion) | +$0.1142 ± 0.0098 | +$0.1258 ± 0.0169 |
| LUT teacher, flat stake | +$0.0627 ± 0.0058 | +$0.0770 ± 0.0095 |
| This model, own stake | +$0.0201 ± 0.0091 | +$0.0161 ± 0.0148 |
| This model, edge stake (decode-time) | +$0.0213 ± 0.0155 | +$0.0355 ± 0.0265 |
| No-trade baseline | $0.00 | $0.00 |
| LightGBM / LogReg / DQN / LSTM / LinUCB / Thompson | negative | negative |
- This model beats no-trade pooled across windows (≈ +0.021/game,z≈2.7)butthesimplelookup−tableteacherstaysclearlyahead(+0.11–0.13/game with edge staking).
- Edge-proportional staking applied at decode time more than doubles this model's test
P&L (+0.016→+0.036) with no retraining; val transfer is weaker (+$0.0213).
- Parse-fail rate ≈ 9% (intrinsic, not truncation; invalid outputs default to no-trade).
- Fill realism not modeled (no book depth/latency) — treat as upper bounds.
GRPO v2b (stake-scaled reward) — negative result
Six documented attempts to retrain GRPO with edge-proportional stake in the reward
(script).
The best run trained to ~step 337/500 before collapse: frac_reward_zero_std → 1.0,
grad_norm → 0, ~98% of completions truncated at max length, zero advantage signal.
Cancelled per pre-agreed kill criteria. The v2b adapter is not usable. The
stake-alignment hypothesis (aligning the RL reward with the winning staking scheme)
remains untested end-to-end: TRL GRPO on this 0.6B model with this degenerate-prone
reward structure did not train stably at any explored configuration.
Usage
from peft import AutoPeftModelForCausalLMmodel = AutoPeftModelForCausalLM.from_pretrained( "sach0312/qwen3-0.6b-nhl-polymarket-grpo-v2", torch_dtype="bfloat16")
Prompt format: the sft_v2 state schema (see
sach0312/nhl-polymarket-trading),
response = JSON {"action": "buy_team0|buy_team1|unwind|no_trade", "stake": 0..1, "reason": "..."}.
Training
- TRL
GRPOTrainer, 500 steps, 8 generations/state, temp 1.0, adaptive entropy
(target 1.2), lr 1e-5, a10g-small (~62 min)
- Reward: fee-net counterfactual P&L per action vs game settlement; malformed −0.05,
buy at price >0.95 −0.05, invalid unwind −0.02
- Dashboard: sach0312/qwen3-nhl-grpo-v2-trackio
Limitations
- Research prototype; not financial advice; Polymarket NHL markets are thin
- Price-only settlement replay, 1-minute candles, no order-book depth
- Policy hedges via unwind only; no partial fills or latency modeling