Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
Loading...
If you have background in AI, ML, data science or statistics (and these overlap!) then you’re well suited to become a predictoor to make $.
Typical steps as a Predictoor:
Play with predictoor.ai. Go to predictoor.ai to build intuition: observe the free feed, perhaps buy a few feeds, and watch them change over time.
Run a predictoor bot. Follow the steps in the Predictoor README. You’ll start by running simulations with AI-powered predictions. Then you'll run a predictoor bot on a remote testnet staking fake OCEAN. Finally, you’ll do it on mainnet staking real OCEAN.
Optimize the bot. Improve model prediction accuracy via more data and better algorithms. Extend to predict >1 prediction feeds (Predictoor has many). Wash, rinse, repeat.
The actions as a predictoor give the following ways to earn:
Feed sales. At an epoch, sales revenue (minus fees) for that epoch goes to predictoors. It’s distributed pro-rata by stake among the predictoors who predicted the true value correctly. The revenue for an epoch is the fraction of sales, spread uniformly across subscription length. A price of 3 OCEAN, 5m epochs, and 24h (1440m) subscriptions gives a revenue of (# subscribers) * (3 OCEAN) * / (1440m / 5m).
Stake reshuffling. At an epoch, incorrect predictoors have their stake slashed. This slashed stake is distributed to the correct predictoors pro-rata on their stake.
Predictoor Data Farming. This amounts to additional earning for predictoors.
On Prediction Accuracy. Don’t expect to be 100% accurate in your up/down predictions. Marginally better than 50% might be enough, and be skeptical if you’re greatly above 50%, you probably have a bug in your testing.
On Amount to Stake. How much should a predictoor stake, to maximize revenue and bound risk? If you stake more, you can earn more, but only to a point. Earnings are bounded by sales of the feeds. So if you stake too much, you will lose $. provides equations that bound how much to stake, along with general tips.
You will lose money as a predictoor if your $ out exceeds your $ in. If you have low accuracy you’ll have your stake slashed a lot. Do account for gas fees, compute costs, and more. Everything you do is your responsibility, at your discretion. None of this blog is financial advice.
"Tomorrow belongs to those who can hear it coming." — David Bowie
, artificial or otherwise. We dream of a world of 10,000 truly accurate prediction feeds, for everything from rain forecasts to sea level rise, or traffic congestion to ETH price. is an on-chain, privacy-enabled, AI-powered application and stack that is bringing this dream to reality.
Accurate predictions are valuable. With them, one can take action and create value. Conversely, inaccurate predictions lead to disaster. Predictions have value because they're the , right before action is taken by the user.
Prediction feeds are a stream of predictions for a given time series. This could be predicting the price of ETH every 5 minutes, or the sea temperature daily. A feed may be binary, i.e. whether a time series changes up or down: ↑↓↓↓↑↓↑↑. Accurate prediction feeds are valuable.
Alas, accurate predictions are hard. Worse, typical prediction feeds have no accountability on accuracy. If the weatherman says "no rain for today" and then it rains, a farmer could get stuck in the mud, wrecking a portion of his crops. The weatherman doesn't feel the impact of wrong predictions, but the farmers sure care!
Accountable Predictions - Imagine if there was accountability. Accuracy would go up; the farmer would be stuck less. Imagine accountable prediction feeds for not only for rain, but also wind, sea temperature, road congestion, train delays, ETH prices, NVID prices, housing prices, and more. Imagine tens of thousands of prediction feeds with accountable accuracy. Imagine them globally distributed, and censorship resistant. Imagine accuracy improving with time.
Ocean Predictoor is a stack and a dapp for prediction feeds. It has accountability for accuracy, via staking. It’s globally distributed and censorship-resistant, by being on-chain. We expect its accuracy to improve over time, due to its incentive structure. Its first use case is DeFi token prediction because users can close the data value-creation loop quickly to make tangible $.
Prediction feeds are crowd-sourced. "Predictoor" agents submit individual predictions and stake on them. They make money when they're correct and lose money when not. This drives accurate prediction feeds, because only accurate predictoors will be making $ and sticking around.
“Trader” agents buy aggregate predictions, then use them to take action like buying or selling. The more accurate the predictions, the more easily they make $, the longer they stick around to keep buying prediction feeds from trading profits.
Predictoor is built on the Ocean Protocol stack, including contracts for tokenized data and middleware to cache metadata. To keep predictions private unless paid for, Predictoor uses Oasis Sapphire privacy-preserving EVM chain.
The initial dapp is live at predictoor.ai. It’s for up/down predictions of BTC, ETH, and other tokens’ prices. The dapp helps users build a mental model of Predictoor behavior. Predictoors and traders’ main workflow is to do run predicting / trading bots with the help of the Py SDK. We have seeded Predictoor with bots that have AI/ML models of accuracy comfortably above 50% — a precondition to make $ trading.
Run AI-powered prediction bots or trading bots on crypto price feeds to earn $
Ocean Predictoor provides on-chain "prediction feeds" on whether ETH, BTC, etc will rise in the next 5 min or 60 min. "Predictoors" submit predictions and stake on them; predictions are aggregated and sold to traders as alpha. Get started at predictoor.ai.
- the "why" and "what"
- structure, behavior, privacy
How to earn as a ...
- running a prediction bot
- running a trading bot
- baseline sales $ for running predictoor bots
on price of feeds, more
and
Core repo:
Finally, join us on to chat with other predictoors, traders, or devs in the Predictoor ecosystem.
Tomorrow belongs to those who can hear it coming. — David Bowie
Data Farming (DF) is Ocean's incentives program. There are weekly OCEAN rewards for various activities.
The main Predictoor-based way to earn from DF is to become a predictoor, and earn directly from Predictoor DF.
Main:
Mainnet webapp: predictoor.ai. Testnet: test.predictoor.ai
Run bots via pdr-backend repo
Intro:
While these docs are the best starting point to learn about Predictoor, here are some other useful sources.
Original blogpost "Meet Predictoor", Sep 12, 2023 []
Original talk "Introducing Prediction Feeds", Dappcon, Sep 12, 2023 [][][]
For thoroughness:
All blog posts & media:
Webapp code is at repo
Congrats, you've gone through all the Predictoor docs!
Next: Go earn! Start at . Or, go to
You can make $ by buying prediction feeds, and using it as an input — as “alpha” — to your trading approach.
Typical steps as a Trader:
Play with dapp, and trade. First, go to to build intuition: observe the free feed, perhaps buy a few feeds, and watch them change over time. In a second window, have Binance open. Employ a baseline trading strategy: when a new Predictoor prediction pops in, buy if “↑”, and sell or short if “↓”; exit the position 5 min later.
Run a trader bot. Follow the steps in the . You’ll start by running simulations with AI-powered predictions. It follows the baseline trading strategy: when a new Predictoor prediction pops in, it buys if “↑”, and sells if “↓”. Then you’ll do run a trader bot on a remote testnet with fake tokens. Finally, you’ll do it on mainnet with real tokens on a real exchange.
Improve & extend. Improve trading performance via more sophisticated trading strategies. This is a universe all of its own! Extend to >1 prediction feeds (Predictoor has many). Wash, rinse, repeat.
The actions as a trader offer a single yet powerful way to earn: trading revenue. Buy low and sell high! (And the opposite with shorting)
⚠️ You will lose money trading if your $ out exceeds your $ in. Do account for trading fees, order book slippage, cost of prediction feeds, and more. Everything you do is your responsibility, at your discretion. None of this blog is financial advice.
Predictoor aims to make it easy for people to make $ doing predictions (as predictoors), and taking actions against those predictions (as traders).
Before we started building Predictoor, we first asked: can we predict ETH (etc) up/down with accuracy? Then we conducted intensive AI/ML research towards this question. The results of this investigation were positive 😎. TBH, we were surprised at the degree of market inefficiency.
Our next question was: with these predictions, can we make $ trading? Again, we conducted intensive AI/ML research, and again found positive results 😎😎.
Here we share some results of that research — a glimpse of how deep the rabbit hole goes — to inspire would-be predictoors and traders in their own work. The model was trained on data from January 1, 2021 to June 30, 2023, with simulated results the first 24 days of July 2023. A “baseline” trading strategy was used:
The image below shows simulated returns as a function of order size, for BTC/USDT on Binance. Duration = 7000 ticks x 5m/tick = 24.3 days of trading. It simulates spread effects. It assumes 0% fees. Note how the size of the order affects the return. This is because BTC/USDT is not very liquid; therefore larger amounts cause slippage.
Simulated returns vs time of BTC/USDT trading on Binance. Trade size has an impact. The image below has the same experimental setup, but for BTC/USDT pair. The size of the order does not affect the return, because BTC/USDT is more liquid than BTC/USDT.
"It was never easy to look into the future, but it is possible and we should not miss our chance." — Andrei Linde
Contents:
The image below gives an overview of Predictoor structure.
In the image top left, predictoors, traders, or anyone play with predictoor.ai to build an understanding how predictoor works. One feed is free; the rest are available for purchase. At first, only the free feed is visible. Users can connect their web3 wallet and buy another feed.
In the image top middle, predictoors graduate to building & deploying Python "Template Predictoor bots" (agents), which submit predictions every 5 minutes. Now, predictoors can see how to make revenue from making predictions, with plenty of room to improve AI/ML modeling accuracy and make more $.
In the image top right, traders graduate from predictoor.ai to building & deploying Python "Template Trader bots" (agents), which grab the latest prediction every 5 minutes, as soon as it's available, then trade using that prediction (and other info). Now, Traders can see how to make $ from buying predictions, with plenty of room to improve trading strategy and make more $.
In the image bottom is the Oasis Sapphire chain, with Predictoor feed contracts deployed to it. There's one contract deployed for each {pair, exchange, timescale} such as {ETH/USDT, Binance, 5m}.
We just covered Predictoor structure. Let's now layer on some Predictoor behavior, with the help of the image below. We'll walk through actions by Predictoors and Traders related to predictions for time slot "epoch t+1", and show how they make $.
We assume predictions on BTC, and where epoch t ends at 5:00pm, t+1 ends 5:05pm, and t+2 ends 5:10pm. We assume that Traders already purchased a subscription via predictoor.ai, Python, or otherwise. When we discuss an action by a Predictoor or Trader, we recognize that it's typically executed by their agent (bot).
Epoch t. This is left 1/3 of the image. It starts at 4:55pm and ends at 5:00pm. Predictoor 1 (pink) predicts that BTC close price for epoch t+1 will be higher ("↑") than close price for epoch t. He submits a tx to that chain with that prediction, and some OCEAN stake of his choice (higher stake = more confident). Predictoor 2 (dark green) does the same. Predictoor 3 (light green) predicts "↓" and stakes. The chain stores these prediction values, privately.
Epoch t+1. The middle 1/3 of the image covers epoch t+1. It starts at 5:00pm and ends at 5:05pm. The BTC Predictoor contract computes the aggregated predicted value (agg_predval) as stake-weighted sum across individual predictions.
agg_predval = (stake1 x predval1 + stake2 x predval2 + …) / (stake1 + stake2 + …)
The contract then makes agg_predval visible to its subscribers. The predicted value is the stake-weighted sum across predictions. Smart traders may take the information and act immediately. A baseline strategy is "if it predicts ↑ then buy; if it predicts ↓ then sell or short".
Epoch t+2. This is the right 1/3 of the image. It starts at 5:05pm and ends at 5:10pm. Both traders and trueval agent take action (and, predictoors get paid).
Actions by Traders. Typically, traders exit their position immediately, exactly 5 minutes since they got the 5-minute-ahead prediction and acted*. If the prediction feed was accurate enough and trading fees & slippage weren't too high, then the trader makes money on average.
Actions by Trueval agent; predictoors get paid. The trueval agent is a process that grabs price feeds from e.g. Binance and submits it to chain, inside the smart contract*. The "submit" transaction also takes the opportunity to calculate each Predictoor's reward or slashing, and update their OCEAN holdings in the contract accordingly. (Funds aren't sent per se, they're made available via ERC20 "approve", for the predictoor to transfer at some later point). Predictoor 3 got his OCEAN slashed because he was wrong; predictoors 1 and 2 receive that as income, in addition to receiving income from prediction feed sales to traders. Predictoors can claim their balance anytime.
Predictoor needs privacy for:
Submitted predictions
Compute aggregate predictions; and
Aggregated predictions — only subscribers can see
This could all be done on fully-centralized infrastructure. But doing so would fail on our other goals: being globally distributed, censorship resistant, and non-custodial.
Targeting these needs, we researched & prototyped many privacy technologies. emerged as the best choice because, as the only privacy-preserving EVM chain in production, it could handle these needs cleanly end-to-end.
Most of dapp is implemented in the , with help from to fetch feed data and for metadata caching.
The template Predictoor bots, trader bots, trueval bot, and prediction feed publishing are all in the . Contracts are in the Ocean repo.
Events emitted by contracts are indexed as Ocean subgraphs, to be consumed by predictoor.ai and the bots. The backend has more info.
The "Predictoor Structure" section above presented much of the Predictoor architecture.
The image below adds detail around the backend (bottom 1/3 of diagram). Let’s discuss.
Smart Contracts. There’s one Predictoor contract for each prediction feed, at each exchange/timescale: BTC/USDT at Binance/5m, ETH/USDT at Binance/5m, and so on. Each contract is an Ocean datatoken contract, with a new template for prediction feeds.
The implementation is in templates/ERC20Template3.sol at at . It implements ERC20, Ocean, and Predictoor-specific behavior as follows.
- ERC20 behavior. It implements the ERC20 interface and therefore plays well with ERC20-friendly crypto wallets, DEXes, etc.
- Ocean behavior. Being part of Ocean, having 1.0 datatokens means you can access the underlying data asset for the duration of the subscription (once you’ve initiated the order). For Predictoor contracts this is 24h. Each datatoken contract has a parent Ocean data NFT with metadata, means to specify & collect fees, and more.
- Predictoor behavior. Each datatoken contract has additional methods specific to Predictoor: submitting predictions, submitting truevals, computing aggregated predictions, etc.
As of Oct 10, 2023, Predictoor / Ocean contracts are deployed to Oasis Sapphire mainnet.
Users need (real) OCEAN & ROSE tokens: - OCEAN. Staking & payment is in OCEAN. How to get OCEAN on Sapphire. - ROSE. Gas fees are in ROSE. How to get ROSE on Sapphire.
For testnet, there are 10 feeds: X/USDT pair for each of the top-10 coins by market cap (ignoring stablecoins), 5m timescales, on Binance, >0% fees on Binance. Paid feeds. The coins are: X = BTC, ETH, BNB, XRP, ADA, DOGE, SOL, LTC, TRX, DOT
For both testnet and mainnet, there are 20 feeds:
X/USDT pair for each of the top-10 coins by market cap (ignoring stablecoins), on Binance, 5m timescales, on Binance. The coins are: X = BTC, ETH, BNB, XRP, ADA, DOGE, SOL, LTC, TRX, DOT
Like above, but for 60m timescales
For each timescale, one feed is free: BTC/UTD on Binance. Below is pricing for the remaining feeds.
The price to subscribe to one feed for 24 hours is 3.00 OCEAN. This includes all fees. Fee details:
0.1% community swap fee
20% fee to Ocean Protocol Foundation. (Will be used to further drive Predictoor, and to burn OCEAN.)
For reference, price without fees is 2.49791840133 OCEAN. To calculate this: Let x = price without fees. Then x * (1 + 0.20 + 0.001) = 3.0 → x = 3.0 / (1 + 0.20 + 0.001) = 2.49791840133
Pricing is subject to change based on learnings, and feedback from community.
(a) Core team-supplied Python bot
(b) simple Python script
(c) Use "Write Contract" in Sapphire blockchain explorer
(d) 3rd-party bot that decides to support this
(e) dapp that decides to support this
Numerai focus is tradfi trading, as a hedge fund.
Whereas Predictoor is pure datafeeds, and applies to any vertical. Its first use case is defi trading.
Currently Numerai is mostly centralized.
Whereas Predictoor is decentralized. Predictoor tech may be useful to Numerai to help decentralize
Being decentralized means one doesn't need to rely on a centralized actor running centralized services. Useful for anyone coming to rely on prediction feeds.
Finally, Predictoor's ambition is broader: to extend beyond crypto into energy, weather, agriculture and more. It's easier to expand scope of decentralized feeds, because you don't need to spin up a new team to focus on each new vertical.
