Example requests for dojo

wwklsq commented 1 year ago

Hi Team dojo,

I am a beginner with dojo package and struggled a little bit with creating an example script using dojo to provide liquidity and inspect the fees earned in a backtesting manner. So far, I have set up the environment and successfully executed run.py both in notebook as well as in script. Here are the problems I faced today:

Documentation link is broken on tutorial page under code refererence, I believe the link is outdated, this link might be the official docs for dojo.
Failed to use USDC/ETH 0.05% pool (stated here to be supported) to create trade simulation, got key Error from error message.

The minimal viable strategy I wanted to create(cloned and modified from PassiveLP) is to use pool e.g. USDC/WETH 0.3% to mint a position and we should be able to check the uncollected fees in nft position at the end of simulation. Following code snippet explains my idea also the puzzling parts in comment after inspection of docs.

class Simple_lp_deploy(BasePolicy):
"""Provide liquidity passively to a pool in the sepcified price bounds."""

def __init__(self, agent: BaseAgent, lower_price: float, upper_price: float) -> None:
    """Initialize the policy.

    :param agent: The agent which is using this policy.
    :param lower_price: A custom lower price.
    :param upper_price: A custom upper price.
    """
    super().__init__(agent=agent)
    self.lower_price_bound = Decimal(lower_price)
    self.upper_price_bound = Decimal(upper_price)
    self.has_invested = False

def fit(self):
    pass

def inital_quote(self, obs: UniV3Obs) -> List[UniV3Action]:
    pool_idx = 0
    pool_address = obs.pools[pool_idx]
    token0, token1 = obs.pool_tokens(obs.pools[pool_idx])
    wallet_portfolio = self.agent.erc20_portfolio()

    address0, address1 = obs.pool_token_addresses(obs.pools[pool_idx])
    decimals0 = money.get_decimals(self.agent.backend, address0)
    decimals1 = money.get_decimals(self.agent.backend, address1)

    lower_price_range = self.lower_price_bound
    upper_price_range = self.upper_price_bound
    tick_spacing = obs.tick_spacing(pool_address)

    lower_tick = uniswapV3.price_to_tick(
        lower_price_range, tick_spacing, [decimals0, decimals1]
    )
    upper_tick = uniswapV3.price_to_tick(
        upper_price_range, tick_spacing, [decimals0, decimals1]
    )
    provide_action = UniV3Action(
        agent=self.agent,
        type="quote",
        pool=pool_address,
        quantities=[wallet_portfolio[token0], wallet_portfolio[token1]],
        tick_range=(lower_tick, upper_tick),
    )
    self.has_invested = True
    return [provide_action]

def predict(self, obs: UniV3Obs) -> List[UniV3Action]:
    if not self.has_invested:
        return self.inital_quote(obs)

    # https://readthedocs.compasslabs.ai/dojo.observations.html#dojo.observations.UniV3Obs.lp_quantities
    # https://readthedocs.compasslabs.ai/dojo.observations.html#dojo.observations.UniV3Obs.nft_positions
    # I want to print the uncollected fees in each step for our minted position, where can I get the
    # toked_id for our position?
    # print(obs.nft_positions(token_id=))

    # alternatively, I can also call https://readthedocs.compasslabs.ai/dojo.observations.html#dojo.observations.UniV3Obs.pool_positions
    # but what is the owner address in this case? Should it be a real owner address? Or can I
    # get the synthetic owner address from self.agent?
    # print(obs.pool_positions(pool=pool_address, owner=))
    return []

Last but not least, in which unit is rewards defined? I assume, it´s in USDC for run.py case. Do we have a lookup table of definition for other cases? Btw, When I ran the run.py in notebook and plotted the reward results, there´s a big plummet in one block, which looks not close to the result from homepage. Although the time span was different, it also didn´t match any pattern from chart below.

Thanks for the great work in general, I hope I can get supports for questions above and more examples from compass labs side regarding this project.

Best regards, Wwk

0xprofessooor commented 1 year ago

New version released, some more changes include:

port instance can be set to auto close upon successful simulation run.
'live' backend type for generating signals in real time
more helpful observation objects allowing you to easily get information about uniswap and tokens.

Still looking into the final edge case, will update with progress :)

wwklsq commented 1 year ago

Hi @09tangriro,

Testing with latest version of dojo, I had follow-up questions:

Was the data loading issues solved for local backend mode only or also for anvil mode? I´m testing the edge case with different result between chain data and dojo from previous commit, weird though, previously I can run this position with anvil backend, now it yields STF error, probably because it has more data now it leads to overflow in whale account?
What´s the live backend type for? Should we switch to that to get new data, or it´s more likely for live trading purpose.
Can u point me to the docs where are these new observation objects? UniswapV3Obs doesn´t look much different for me.

0xprofessooor commented 1 year ago

So we really only recommend using the forked backend for testing your code for bugs, the local backend should guarantee zero STF errors. If you want to continue using the forked backed, you can try to fund the market agent even more but you'll need to find the relevant whale accounts for that time period: env.agents[0].fund_erc20("USDC", your_whale_account, Decimal(large_number)), you'll likely need to write your own loop though as well:
```
agent = YourAgent()
policy = YourPolicy(agent)

env = UniV3Env(...)

obs = env.reset()
for block in env.iter_block():
    agent_actions = policy.predict(obs)
    market_actions = env.market_actions(agents_actions=agent_actions)
    actions = agent_actions + market_actions
    _, rewards, _, _ = env.step(actions)
```
The live backend isn't quite an execution envrionment, for now it just allows you to query the observation signals in real time, for example to see what actions your policy will generate to new data.
https://readthedocs.compasslabs.ai/dojo.observations.html#dojo.observations.BaseObs.token_decimals and https://readthedocs.compasslabs.ai/dojo.observations.html#dojo.observations.UniV3Obs.tick_liquidities are the main ones

We've also found a speed improvement 🎉 The first edge case (STF) in the list now runs successfully for me in 282s (not yet in release).

wwklsq commented 1 year ago

Hi @09tangriro,

Thanks for the quick update, I switch now to local backend type.

I disabled dashboard_port and auto_close, as both would lead to a crashdown of my working container. Reasons for that are still unclear, as monitoring memory usage and logging didn´t allude any useful information. Presumably working in container is not default/recommended set-up for dojo. I´m working with a handcrafted port-closing function & mem cleaning procedure.
I can confirm by gluing data, the example above has now closed up the fee with the ground-truth.
Although there are still outliers lingering around where dojo either overpredicts the fee or underpredicts the fee. Here are some samples for that case:

[{'start_time': Timestamp('2023-08-17 07:43:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-08-18 10:34:00+0000', tz='UTC'),
  'deposit_token0': 574.861008,
  'deposit_token1': 574.861008,
  'lower_price': 0.0005261,
  'upper_price': 0.00058962,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 3000,
  'dojo_fees_collected_token0': '2.804471',
  'real_fees_collected_token0': 0.581423,
  'dojo_fees_collected_token1': '0.001909294938952954',
  'real_fees_collected_token1': 0.0013293,
  'pool_desc': 'USDC/WETH-0.3',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/551745'},
 {'start_time': Timestamp('2023-04-02 20:13:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-04-04 12:13:00+0000', tz='UTC'),
  'deposit_token0': 39539.1375,
  'deposit_token1': 39539.1375,
  'lower_price': 0.00053565,
  'upper_price': 0.00058962,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 3000,
  'dojo_fees_collected_token0': '304.511513',
  'real_fees_collected_token0': 152.703298,
  'dojo_fees_collected_token1': '0.104557025810633318',
  'real_fees_collected_token1': 0.05234959,
  'pool_desc': 'USDC/WETH-0.3',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/481049'},
 {'start_time': Timestamp('2023-09-15 00:26:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-16 01:34:00+0000', tz='UTC'),
  'deposit_token0': 71236.007076,
  'deposit_token1': 71236.007076,
  'lower_price': 0.00061491,
  'upper_price': 0.00062608,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '180.910437',
  'real_fees_collected_token0': 97.614772,
  'dojo_fees_collected_token1': '0.116631441591324158',
  'real_fees_collected_token1': 0.05905205,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/564394'},
 {'start_time': Timestamp('2023-09-18 13:20:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-19 15:02:00+0000', tz='UTC'),
  'deposit_token0': 2653.713952,
  'deposit_token1': 2653.713952,
  'lower_price': 0.00057506,
  'upper_price': 0.00060576,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '16.987608',
  'real_fees_collected_token0': 7.21048,
  'dojo_fees_collected_token1': '0.010694034637289168',
  'real_fees_collected_token1': 0.00515498,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/565818'},
 {'start_time': Timestamp('2023-09-14 23:57:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-16 06:53:00+0000', tz='UTC'),
  'deposit_token0': 21032.118402,
  'deposit_token1': 21032.118402,
  'lower_price': 0.00060274,
  'upper_price': 0.00061491,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '941.429875',
  'real_fees_collected_token0': 4282.51329,
  'dojo_fees_collected_token1': '0.584749281867812615',
  'real_fees_collected_token1': 2.44181838,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': '[https://revert.finance/#/unis](https://revert.finance/#/uniswap-position/mainnet/564389)
{'start_time': Timestamp('2023-09-11 12:17:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-12 21:49:00+0000', tz='UTC'),
  'deposit_token0': 10153.796168,
  'deposit_token1': 10153.796168,
  'lower_price': 0.00062859,
  'upper_price': 0.00072958,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '1.013681',
  'real_fees_collected_token0': 20.268507,
  'dojo_fees_collected_token1': '0.000594838674044158',
  'real_fees_collected_token1': 0.01295848,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/562977'},
 {'start_time': Timestamp('2023-09-27 11:37:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-28 21:40:00+0000', tz='UTC'),
  'deposit_token0': 9266.97624,
  'deposit_token1': 9266.97624,
  'lower_price': 0.00061001,
  'upper_price': 0.00062047,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '83.143482',
  'real_fees_collected_token0': 53.968819,
  'dojo_fees_collected_token1': '0.050544862935275553',
  'real_fees_collected_token1': 0.03040056,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/570290'},
 {'start_time': Timestamp('2023-09-06 06:10:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-07 18:05:00+0000', tz='UTC'),
  'deposit_token0': 45300.0,
  'deposit_token1': 45300.0,
  'lower_price': 0.00058844,
  'upper_price': 0.00061491,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '1012.024289',
  'real_fees_collected_token0': 2575.259249,
  'dojo_fees_collected_token1': '0.633466098131261847',
  'real_fees_collected_token1': 1.4826215,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/560570'},
 {'start_time': Timestamp('2023-09-10 18:06:00+0000', tz='UTC'),
  'end_time': Timestamp('2023-09-12 09:49:00+0000', tz='UTC'),
  'deposit_token0': 2040.61717,
  'deposit_token1': 2040.61717,
  'lower_price': 0.00061062,
  'upper_price': 0.00062922,
  'token0': 'USDC',
  'token1': 'WETH',
  'fee_tier': 500,
  'dojo_fees_collected_token0': '24.323339',
  'real_fees_collected_token0': 7.348751,
  'dojo_fees_collected_token1': '0.014223481288578276',
  'real_fees_collected_token1': 0.00521725,
  'pool_desc': 'USDC/WETH-0.05',
  'position_nft': 'https://revert.finance/#/uniswap-position/mainnet/562571'}]

0xprofessooor commented 1 year ago

Cool, we'll have a look :)

0xprofessooor commented 1 year ago

Interesting, so the forked backend recovers the true result:

https://revert.finance/#/uniswap-position/mainnet/551745 Screenshot 2023-10-23 at 11 44 47

But the local backend gives incorrect fees as you mention. I'll continue debugging it but in the meantime, for these smaller simulation lengths, the forked backend should suffice.

wwklsq commented 1 year ago

Just noticed, the deposit_token1 is identical to deposit_token0 in my previous commit. It´s a typo. U can probably find the information in the link though.

0xprofessooor commented 1 year ago

Yeah nw, I've been using the link like you said :)

So quick update, I've identified why the local backend was diverging and think I've implemented a fix, am testing now but will update here when I have it working

0xprofessooor commented 1 year ago

Ok so I think that's a confirmed fix now 🎉, will release a new version soon. This will have quite a few changes actually, the interface you will use has also changes so you use UniV3Trade, UniV3Quote etc. rather than UniV3Action, which I hope is a bit more intuitive. I'll do the release soon but updating the docs at dojo.compasslabs.ai might take a bit longer.

wwklsq commented 1 year ago

Ok so I think that's a confirmed fix now 🎉, will release a new version soon. This will have quite a few changes actually, the interface you will use has also changes so you use UniV3Trade, UniV3Quote etc. rather than UniV3Action, which I hope is a bit more intuitive. I'll do the release soon but updating the docs at dojo.compasslabs.ai might take a bit longer.

Hi @09tangriro , thanks for the quick fix. Can I assume that dojo has now fixed all the problematic positions that I listed above, so the result from dojo in local mode now matches exactly with the groundtruth from revert?

0xprofessooor commented 1 year ago

I believe so, just did a release now, you'll need to edit your code slightly to use the new format for actions though (https://readthedocs.compasslabs.ai/dojo.actions.html#dojo.actions.uniswapV3.UniV3Quote)

0xprofessooor commented 1 year ago

Thanks for giving dojo a try, closing the issue now :)

CompassLabs / dojo_examples

Example requests for dojo #15