Internal rules for deciding whether the dashboard is finding tradable edge.
# Trading Operating Model
Prediction Alpha Desk is a research and paper-trading system for finding public-information mispricings in prediction markets.
The purpose is not to generate interesting market commentary. The purpose is to determine whether a repeatable, scalable process can find enough high-edge markets to justify eventual real-money deployment.
This document defines the operating rules for v0/v1.
## Core question
For every market, ask:
> Can we estimate fair probability materially better than the market, using public information, resolution mechanics, deadlines, legal/procedural bottlenecks, local-language information, or cross-market inconsistencies?
If the answer is not clearly yes, the decision is WATCH or SKIP.
## Current strategic shift
The California billionaire-tax ballot trade showed that Prediction Alpha Desk should not only think in terms of holding paper bets to resolution.
The more powerful model may be:
> stale public-information probability -> paper entry -> market repricing -> simulated exit -> redeploy
That means the project should evaluate both:
1. **settlement accuracy** — did the market ultimately resolve the way our thesis expected?;
2. **repricing monetization** — did the market move enough in our favor that a real trader could plausibly have exited profitably before settlement?
This is still paper only. The goal is to learn which entry/exit/sizing rules would have worked before risking real money.
See also: `docs/SIMULATION_AND_EXIT_MODEL.md`.
## System layers
Prediction Alpha Desk has four layers:
1. Screener
- Finds potentially interesting markets.
- Uses deterministic heuristics.
- Does not decide trades.
2. Research funnel
- Converts candidates into SKIP, WATCH, PAPER YES, or PAPER NO.
- Requires explicit reasoning and source-backed probability estimates.
3. Paper portfolio simulator
- Tracks paper trades, sizing, open P/L, resolved outcomes, and postmortems.
- Tracks both mark-to-market and settlement outcomes.
4. Scenario optimizer
- Tests which entry buckets, confidence levels, edge thresholds, holding periods, and exit rules would have produced the best paper results.
- This layer is analytical only. It does not trade.
## Market decisions
### SKIP
Use SKIP when any of the following applies:
- Resolution criteria are unclear or too subjective.
- The market depends mostly on vibes, narrative, celebrity attention, or unknowable private decisions.
- Liquidity is too weak for a realistic entry.
- The price may be directionally wrong but the expected ROI is too low for the current test bucket.
- The evidence chain is too long relative to the likely edge.
- The market is interesting but not tradable.
### WATCH
Use WATCH when the market is interesting but one or more key conditions are missing:
- One critical fact needs confirmation.
- The price is not yet good enough.
- A deadline, official document, court/agency step, or event calendar is pending.
- The market belongs to a cluster that needs comparison before taking a side.
WATCH is not a trade.
### PAPER YES / PAPER NO
Use PAPER YES or PAPER NO when the trade is credible enough to be useful for the simulator.
There are two standards:
1. **High-conviction paper trade**
- This is the old standard: it would be tempting even with real money later.
- Use larger paper sizing.
2. **Learning paper trade**
- This is allowed in paper mode when the market falls into a strategically important test bucket.
- Use smaller paper sizing.
- Must still have a clear thesis and measurable fair-value gap.
- Must not be random action for the sake of action.
Minimum requirements for any PAPER trade:
- Resolution criteria are clear.
- The thesis can be stated in one or two precise sentences.
- The estimated fair probability differs from market price by enough to test a defined bucket.
- Liquidity and spread are good enough for a realistic paper entry.
- The trade is not merely a vague narrative call.
A strong paper trade usually has:
- 10+ percentage-point edge.
- 15%+ expected ROI.
- Clear public evidence.
- Short or medium evidence chain.
- Limited dependency on private information.
Do not require every paper trade to have California-level upside. A 50% market with credible 75% fair value may be a valid and important paper trade because the portfolio may prefer higher hit-rate, lower-ROI opportunities.
## Probability estimates
Use probability bands when precision would be fake.
Suggested labels:
- Very unlikely: 0-10%
- Unlikely: 10-30%
- Toss-up: 40-60%
- Likely: 70-90%
- Very likely: 90%+
Always record low, mid, and high fair-probability estimates where possible.
## Expected value and ROI
For a YES position:
- Edge per share = fair YES probability - YES price.
- Expected ROI = edge per share / YES price.
For a NO position:
- NO price = 1 - YES price, unless the market gives a separate NO price.
- Fair NO probability = 1 - fair YES probability.
- Edge per share = fair NO probability - NO price.
- Expected ROI = edge per share / NO price.
Do not confuse win probability with expected value. A high win-rate strategy can still be bad if one loss destroys many small gains.
## Entry buckets to test
Do not prejudge the ideal price/conviction mix. Track it.
Suggested buckets:
| Bucket | Example | Why it matters |
|---|---|---|
| Low-price asymmetric | 5% -> 15% | High ROI, high zero-risk. |
| Medium-price mispricing | 25% -> 45% | Often good upside/survivability balance. |
| High-conviction moderate edge | 50% -> 75% | Lower ROI, potentially higher hit rate. |
| High-price defensive | 80% -> 93% | Low headline ROI, but useful if uncertainty is fake. |
| NO-side fade | YES overpriced, NO underpriced | Captures public overbelief in narrative YES. |
The purpose of paper mode is to find which buckets actually make money.
## Paper sizing
Default paper stakes:
| Confidence | Paper stake |
|---|---:|
| Low / learning bucket | $100 |
| Medium | $250 |
| High | $500 |
| Very high | $1,000 |
Sizing rules:
- Never paper-size beyond what the liquidity could plausibly support.
- As a rule of thumb, paper stake should usually be no more than 1-3% of visible liquidity.
- Markets with extremely low liquidity may be useful for learning but should not be treated as realistic scalable opportunities.
- Avoid very extreme prices unless the edge is unusually strong or the settlement date is close.
- Paper sizing should deliberately test portfolio behavior, not only individual conviction.
## Exit-review model
The current monitor can flag an open PAPER trade as `EXIT_REVIEW`.
This is not a forced exit. It means the position has changed enough that a simulated take-profit decision should be considered.
Current triggers:
- current price is at least 2x entry;
- current price is at least 80% of fair value;
- open P/L is at least +75% of stake;
- profitable position is inside the 7-day deadline window;
- remaining upside to fair is 25% or less.
These are exit triggers, not entry requirements.
## Exit policies to simulate
The simulator should eventually compare multiple exit policies:
1. Hold to resolution.
2. Full exit at first exit-review trigger.
3. Sell enough to recover initial stake and let the rest run.
4. Sell 75%, keep 25% runner.
5. Exit at 80%, 90%, or 95% of estimated fair value.
6. Exit profitable positions inside final 7 days unless outcome is effectively locked.
The goal is to identify the exit policy with the best balance of P/L, drawdown, time-in-trade, and liquidity realism.
## Research format
Each researched market should be logged, including SKIP and WATCH decisions.
Minimum fields:
- Title
- URL
- Current YES price
- Current NO price
- Deadline
- Volume
- Liquidity
- Literal resolution criteria
- Casual-trader misunderstanding
- Timeline and bottlenecks
- Evidence summary
- Fair YES probability range
- Edge versus market
- Decision
- Confidence
- Paper side and stake, if any
- Screener lesson
- Exit notes / scenario notes, once implemented
## Batch workflow
Do not deeply research every candidate.
For each dashboard run:
1. Review top candidates.
2. Select the highest edge-per-research-hour candidates.
3. Include a small number of learning-bucket trades when they test a valuable hypothesis.
4. Log all decisions.
5. Update screener rules only when a repeated pattern appears.
The goal is disciplined conversion from candidate to trade/no-trade while building enough paper critical mass to evaluate the strategy.
## Preferred market features
Prioritize markets with:
- Clear resolution criteria.
- Near or medium-term deadlines.
- Official sources.
- Legal/procedural gates.
- Ballot access, court, agency, parliamentary, party-rule, or election-calendar mechanics.
- Local-language information gaps.
- Cross-market or cluster inconsistencies.
- Prices that imply casual misunderstanding.
## Lower-priority market features
Do not automatically reject these, but demand better evidence:
- Long-dated political popularity markets.
- Celebrity or narrative-driven markets.
- Awards markets with opaque juries.
- Very high-price or very low-price markets with poor capital efficiency.
- Markets where the only thesis is polling, vibes, or general political intuition.
- Markets requiring very deep specialist expertise without a clear public-information trap.
## Candidate-flow stance
Current candidate flow may be too thin.
The answer is not to lower standards randomly. The answer is to broaden toward more California-shaped opportunities:
- ballot access / certification;
- filing deadlines;
- court rulings and injunctions;
- agency approvals;
- legislation passed/signed;
- parliamentary procedure;
- official database/status changes;
- source-specific map/database markets;
- cross-market discrepancies.
Kalshi may be added later as Scout Mode, not immediate full integration.
## Success thresholds
After 25 researched paper decisions:
- Check whether at least 5 credible PAPER trades were found.
- Check whether the average stated edge on PAPER trades was 10+ percentage points.
- Check whether the process is producing actionable candidates often enough.
- Check whether at least one exit policy appears promising.
After 50 researched paper decisions:
- Evaluate paper P/L by settlement and by simulated exit policy.
- Evaluate realized ROI.
- Evaluate win rate by price bucket.
- Evaluate profit per day of capital exposure.
- Evaluate whether wins were for the stated reasons.
- Identify categories that deserve promotion or demotion.
After 100 researched paper decisions:
- Decide whether the project has a real edge.
- If not, narrow the scope or pull the plug.
- If yes, consider a tiny real-money pilot with strict limits.
## Pull-the-plug criteria
Consider stopping or radically narrowing the project if:
- Fewer than 10% of researched candidates become credible PAPER trades.
- Paper P/L is negative after a meaningful sample under realistic exit assumptions.
- Most wins are luck or vague narrative calls.
- The best opportunities are too illiquid to matter.
- The process requires too much manual research for too little expected value.
- The screener repeatedly surfaces markets that cannot be converted into high-confidence decisions.
## Scale criteria
Only consider real-money testing after:
- A meaningful paper ledger exists.
- The process shows positive paper P/L under realistic simulated exit rules.
- The best categories are identifiable.
- The same kind of edge appears repeatedly.
- The losses are understandable and survivable.
- The system has shown discipline in skipping weak markets.
- Liquidity/spread assumptions are realistic.
## Current stance
The project is in paper-trading calibration and simulation mode.
No real-money trading.
No automatic trading.
No hidden discretionary bets.
Do not expand randomly into noise.
Do not deliberately slow learning when paper mode can safely test scenarios.
Every market is either:
- A paper trade,
- A watch item,
- A skip,
- A screener lesson,
- Or a scenario-calibration data point.