Fix #137: pdr dynamic modeling as approach3 (PR #223)

Co-authored-by: trizin <25263018+trizin@users.noreply.github.com>

.gitignore

@@ -163,3 +163,7 @@ cython_debug/
 #.idea/
 .test_cache/
 .cache/
+
+# predictoor dynamic modeling
+out*.txt
+csvs/

READMEs/localpredictoor-localnet.md

@@ -14,10 +14,13 @@ This README describes:
 1. **Setup & run local network**
     - [Install it](#install-local-network)
     - [Run it](#run-local-network)
-2. **Setup & run predictoor bot**
+2. **Setup predictoor bot**
     - [Install it](#install-predictoor-bot)
     - [Set envvars](#set-envvars)
-    - [Run it](#run-predictoor-bot): [random](#run-random-predictoor) then [model-based](#run-model-based-predictoor)
+3. **[Run predictoor bot](#run-predictoor-bot)**
+    - [Random](#run-random-predictoor)
+    - [Static model](#run-static-model-predictoor)
+    - [Dynamic model](#run-dynamic-model-predictoor)
 
 ## Install Local Network
 
@@ -66,10 +69,6 @@ There are other envvars that you might want to set, such as the owner addresses.
 
 Here, we run a bot that makes random predictions.
 
-- It runs [`predictoor_agent1.py::PredictoorAgent1`](../pdr_backend/predictoor/approach1/predictoor_agent1.py) found in `pdr_backend/predictoor/approach1`
-- It's configured by envvars and [`predictoor_config1.py::PredictoorConfig1`](../pdr_backend/predictoor/approach1/predictoor_config1.py)
-- It predicts according to `PredictoorAgent1:get_prediction()`.
-
 In work console:
 ```console
 # run random predictoor bot
@@ -79,20 +78,23 @@ python pdr_backend/predictoor/main.py 1
 Observe the bots in action:
 - In the barge console: trueval bot submitting (mock random) truevals, trader is (mock) trading, etc
 - In your work console: predictoor bot is submitting (mock random) predictions
+- Query predictoor subgraph for detailed run info. [`subgraph.md`](subgraph.md) has details.
 
-You can query predictoor subgraph for detailed run info. [`subgraph.md`](subgraph.md) has details.
-
+Code structure:
+- It runs [`predictoor_agent1.py::PredictoorAgent1`](../pdr_backend/predictoor/approach1/predictoor_agent1.py) found in `pdr_backend/predictoor/approach1`
+- It's configured by envvars and [`predictoor_config1.py::PredictoorConfig1`](../pdr_backend/predictoor/approach1/predictoor_config1.py)
+- It predicts according to `PredictoorAgent1:get_prediction()`.
 
-### Run Model Based Predictoor
+### Run Static Model Predictoor
 
-Since random predictions aren't accurate, let's use AI/ML models. Here's an example flow that loads pre-learned models ("approach2"):
+Since random predictions aren't accurate, let's use AI/ML models.
+- Here in "approach2", we load pre-learned models (static)
+- And in "approach3" further below, we learn models on-the-fly (dynamic)
 
+Code structure:
 - The bot runs from [`predictoor/approach2/main.py`](../pdr_backend/predictoor/approach2/main.py), using `predict.py` in the same dir.
 - Which imports a model stored in [`pdr-model-simple`](https://github.com/oceanprotocol/pdr-model-simple) repo
 
-
-Once you're familiar with this, you'll want to fork it and run your own.
-
 In work console:
 ```console
 # (ctrl-c existing run)
@@ -113,6 +115,61 @@ pip install scikit-learn ta
 python pdr_backend/predictoor/main.py 2
 ```
 
+Once you're familiar with this, you can fork it and run your own.
+
+### Run Dynamic Model Predictoor
+
+Here, we build models on-the-fly, ie dynamic models. It's "approach3".
+
+Whereas approach2 has model development in a different repo, approach3 has it inside this repo. Accordingly, this flow has two top-level steps:
+1. Develop & backtest models
+2. Run bot in Predictoor context
+
+Let's go through each in turn.
+
+**Approach3 - Step 1: Develop & backtest models**
+
+Here, we develop & backtest the model. The setup is optimized for rapid iterations, by being independent of Barge and Predictoor bots.
+
+In work console:
+```console
+#this dir will hold data fetched from exchanges
+mkdir csvs
+
+#run approach3 unit tests
+pytest pdr_backend/predictoor/approach3
+
+#run approach3 backtest.
+# (Edit the parameters in runtrade.py as desired)
+python pdr_backend/predictoor/approach3/runtrade.py
+```
+
+`runtrade.py` will grab data from exchanges, then simulate one epoch at a time (including building a model, predicting, and trading). When done, it plots accumulated returns vs. time. Besides logging to stdout, it also logs to out*.txt in pwd.
+
+**Approach3 - Step 2: Run bot in Predictoor context**
+
+Once you're satisfied with your backtests, you're ready to run the approach3 bot in a Predictoor context.
+
+First, get Barge & other bots going via "Run local network" instructions above.
+
+Then, in work console:
+```console
+# run approach3 predictoor bot
+python pdr_backend/predictoor/main.py 3
+```
+
+Observe all bots in action:
+- In the barge console: trueval bot submitting (mock random) truevals, trader is (mock) trading, etc
+- In your work console: predictoor bot is submitting (mock random) predictions
+- Query predictoor subgraph for detailed run info. [`subgraph.md`](subgraph.md) has details.
+
+Code structure:
+- It runs [`predictoor_agent3.py::PredictoorAgent3`](../pdr_backend/predictoor/approach3/predictoor_agent3.py) found in `pdr_backend/predictoor/approach3`
+- It's configured by envvars and [`predictoor_config3.py::PredictoorConfig3`](../pdr_backend/predictoor/approach3/predictoor_config3.py)
+- It predicts according to `PredictoorAgent3:get_prediction()`.
+
+Once you're familiar with this, you can fork it and run your own.
+
 ## Next step
 
 You're now running a local predictoor bot on a local network. Congrats!

mypy.ini

@@ -56,5 +56,8 @@ ignore_missing_imports = True
 [mypy-web3.*]
 ignore_missing_imports = True
 
+[mypy-sklearn.*]
+ignore_missing_imports = True
+
 [mypy-pdr_backend.predictoor.examples.*]
 ignore_missing_imports = True

pdr_backend/predictoor/approach3/constants.py

@@ -0,0 +1,9 @@
+import numpy as np
+
+OHLCV_COLS = ["open", "high", "low", "close", "volume"]
+OHLCV_DTYPES = [np.float64] * len(OHLCV_COLS)
+
+TOHLCV_COLS = ["timestamp"] + OHLCV_COLS
+TOHLCV_DTYPES = [np.int64] + OHLCV_DTYPES
+
+MS_PER_EPOCH = 300000  # 300,000 ms in 5 min