Incentive Design 2.0
Bettensor
Goal:
A scoring mechanic that more effectively differentiates miners with effective strategies over a large volume of predictions. We are attempting to select for sports predictors that would have excellent real-world performance.
Framework:
The core framework for the scoring system is a tiered ranking approach, similar to how many video games rank players for online matches. It is designed to keep miners in competition with others of similar skill levels, and provide a progression for truly good sports predictors to earn significant incentive.
Each tier has:
A scoring window (number of days that each score component is measured over),
a minimum wager amount for the scoring window,
a progressively smaller capacity,
and a progressively higher incentive share.
Tiers and Requirements:
Tier 1
Scoring Window: 3 Days
Minimum wager: N/A
Capacity: N/A (Up to the maximum subnet capacity)
Incentive Share: 10% of total miner emissions
Tier 2
Scoring Window: 7 Days
Minimum wager: $4000 over 7 days
Capacity: ~20% of all miners
Incentive Share: 15% of total miner emissions
Tier 3
Scoring Window: 15 Days
Minimum wager: $10,000 over 15 days
Capacity: ~20% of all miners
Incentive Share: 20% of total miner emissions
Tier 4
Scoring Window: 30 Days
Minimum wager: $20,000 over 30 days
Capacity: 10% of all miners
Incentive Share: 25% of total miner emissions
Tier 5
Scoring Window: 45 Days
Minimum wager: $35,000 over 45 days
Capacity: 5% of all miners
Incentive Share: 30% of total miner emissions
Progression through Tiers (Promotion and Demotion):
Miners, upon initial registration (or when we officially move to this new system on mainnet) will always start in tier 1.
In order for a miner to be promoted to a higher tier, they must meet the minimum requirements for that tier and be in the top 50% of miners in their current tier. If the next tier has open slots, they can be promoted immediately upon meeting the requirements.
Now, here’s where it gets a bit more complicated. Since tiers have a maximum capacity, promotion has another requirement once a higher tier is at capacity:
A miner must beat the score of the lowest miner in the next tier, ** as measured over the higher tier’s scoring window **.
If they achieve this, those two miners will swap places.
If there are multiple miners up for promotion in one scoring run, they will be ranked against the corresponding amount of lowest scoring miners in the next tier. The system will attempt to maximize the number of possible swaps, giving priority to higher scores.
As we can see in the below example, even though miners 1 + 2 could take the positions of miners 6 + 7 respectively, we attempt to maximize swaps and prioritize score, so miners 6 + 7 end up safe.
This logic follows all the way up to tier 5, which will have the strictest wager requirements and be the most competitive over the longest time frame.
Demotions
There is only one condition that would trigger a demotion outside of a swap, and that is a miner not meeting the minimum wager requirement for their current tier (i.e, went inactive/stopped making enough predictions). If this happens, here’s how it works:
The miner being demoted is compared against the scores in the lower tier (as measured over the lower tier’s window). As long as the demoted miner beats the lowest score in the next tier down, they will take that last spot and then be re-ranked according to scores in that tier. If they don’t beat any of those scores either, the process is repeated one tier down from that, and so on.
That corresponding lowest-score miner (the one that is beat by the initially demoted miner) is then demoted in a “cascade” reaction following the same demotion logic.
Order of Operations in Tier Management:
After scores are recalculated (which can happen several times per day), the tier management process follows the following order:
Demotions are checked for and performed first
After demotions have (or have not) occurred, then the swap check process happens.
Scores are recalculated and stored for any miners which have changed tiers, over their new tiers scoring window.
Incentive Distribution
Incentives are always distributed according to the tier definitions, with the exception being when a tier has no miners in it yet. This is likely to only happen when we first switch to this new scoring system, or if we at some point require a “tier reset”.
If a tier has no miners, the allocated incentive will be distributed proportionally to the lower tiers. Once a tier does* have a miner, that incentive allocation returns to that tier.
This creates an interesting opportunity for those that are early to progress through the tiers:
The first miner to reach each higher tier will temporarily receive the ENTIRE allocation of that tier’s incentive, i.e. the race is on! The first miner to Tier 5 will temporarily receive 30% of the entire miner incentive pool! Even once tiers fill up, tiers 4+5 will receive 55% of the incentive pool for the top 15% of miners.
Miner rules:
Miners still get “$1000” per day to predict with.
There is now a minimum wager for each prediction: $10
Beyond that, miners can choose how they allocate their daily wager.
Score Metrics:
The final score is a composite of several key metrics:
ROI
The miner’s return on investment from their wagers.
Sortino Ratio
A similar risk measurement to the sharpe ratio, but one that only penalizes downside volatility. This makes more sense as a metric for sports betting strategies, as many strategies rely on infrequent massive upside volatility to be successful over a long time frame. The calculation for this assumes a risk-free rate of 0.
CLV (Closing Line Value)
A common metric in sports betting, the CLV is a measurement of the difference between odds taken on the wager and closing line odds (odds right before game start). This is a good way to see if a bettor can consistently find “mispriced” odds, indicating some type of edge or success in finding +EV bets.
Entropy System (Experimental Feature, in Progress)
This system is implemented to discourage copy trading, incentive farming and encourage early predictions, contrarian strategies, and incentivize more diversity in data, in general.
Each event (game) is initialized with an “entropy score” based on the inferred probabilities from the average of bookmaker opening odds for each outcome.
Each prediction a miner submits is vectorized according to wager size, pick and odds at time of submission.
The system then measures whether a prediction “adds entropy” to the specific event system. If a prediction varies from the system average, it will add entropy. If it falls in line with the masses, it will either add 0 entropy or subtract entropy. As a part of this system, we have an “internal odds” which shift according to predictions, similar to how bookmakers will sometimes shift their odds according to the direction and volume of wagers they receive leading up to a game. The internal odds aren’t used directly for scoring, but they do help measure how new predictions compare to the event mean.
Predictions which add entropy to an event system are assigned a slight score bonus, relative to the amount of entropy added. Predictions which contribute negative entropy to the system receive a slight score penalty.
The sub-components to the entropy score are:
Prediction Timing/ Order Score
Earlier is better. - Subsequent predictions on an event, especially those with an identical wager amount + pick will receive progressively less score.
Uniqueness Score / Contrarian Bonuses
A distance function from the “consensus prediction” - the mean of all predictions on an event. The more distant + contrarian, the higher the bonus.
Historical Uniqueness
An extra bonus for being consistently “unique” over the scoring window.
This system should have several effects:
Discourages copy trading and incentive farming.
Miners that are “early” to a prediction will receive the highest positive entropy bonuses. Any subsequent predictions, especially if they have equal wager amounts and similar timing, will receive progressively less of a bonus, and even a penalty in some cases (when they are too similar to existing predictions).
This should make it difficult for people to run multiple hotkeys with the same exact predictions and “farm incentive”, taking up a lot of top scoring slots.
Incentivizes earlier prediction, which will provide clearer distinction on CLV scores.
Discourages picking favorites every time (or picking the long shot odds every time)
The Entropy System is a work in progress. It will become more complex over time, eventually including things like more detailed historical analysis, strategy/behavior foot printing, and higher penalties for detected copy trading/incentive farming. The end goal is a system which makes it entirely infeasible for people to copy trade or incentive farm.
Composite Score:
The composite score is a weighted combination of all the above metrics, measured over the scoring window of the tier that the miner resides in:
Each day, each score component is normalized between (-1 → 1) with respect to all scores in all tiers. Then, the composite score for the day is calculated as follows,
Currently, the weights are:
ROI: 30%
CLV: 30%
Sortino: 30%
Entropy: 10%
This could change once we finish running final simulations over the weekend, we are attempting to optimize the final parameters for fairness.
Composite score will be recalculated multiple times per day, each time a validator attempts to set weights. This should account for changes due to new game results.
Final Weight Calculation:
After all scoring calculation and tier management has concluded in a scoring run, the final step is to calculate and allocate weights to every miner. Here are the criteria:
Incentives are distributed according to tier criteria, from top to bottom, to any non-empty tiers. If a tier is empty, its incentive is distributed proportionally to the non-empty tiers.
Bounds are calculated such that the lowest scoring miner in a higher tier always receives at least 1% more incentive than the highest scoring miner in a lower tier.
Within those bounds for a specific tier, incentive is distributed proportionally to scores in the tier.
Finally, weights are normalized to sum to 1 across all miners and submitted to the chain.