RetroChainer

AI engineer · AI agents · Automations.

@CELE8EX

[RetroChainer @RetroChainer]

JUMP TRADING PAYS A QUANT RESEARCHER A MEDIAN $280K A YEAR. THIS ETH DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

lightgbm, adk, pytest and prometheus.

123k stars between them, and the brain is google's

for $0 you get:

→ lightgbm: fast, distributed gradient boosting on decision trees

→ adk: a python toolkit for building, evaluating and deploying ai agents

→ pytest: makes small tests easy to write, and scales to complex ones

→ prometheus: a monitoring system and time series database

how they fit:

1. lightgbm is the eyes: the close landed inside its 90% band on 25 of 30 days

2. adk is the brain: 5 votes, bull vs bear

3. pytest is the judge: 2 tests, the calls vs just holding eth, after costs

4. prometheus is the memory: 1,000 candles, 31 calls and 6 fills

run it yourself:

1 → pip install lightgbm google-adk pytest

2 → one call a day at the 17:00 utc close, logged to prometheus

3 → the boring part: both tests must pass. the calls have to beat holding eth after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have fired the brain in week one. 25 sell calls in 31 days looked broken to me.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

JUMP TRADING PAYS A QUANT RESEARCHER A MEDIAN $280K A YEAR. THIS ETH DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

lightgbm, adk, pytest and prometheus.

123k stars between them, and the brain is google's

for $0 you get:

→ lightgbm: fast, distributed gradient boosting on decision trees

→ adk: a python toolkit for building, evaluating and deploying ai agents

→ pytest: makes small tests easy to write, and scales to complex ones

→ prometheus: a monitoring system and time series database

how they fit:

1. lightgbm is the eyes: the close landed inside its 90% band on 25 of 30 days

2. adk is the brain: 5 votes, bull vs bear

3. pytest is the judge: 2 tests, the calls vs just holding eth, after costs

4. prometheus is the memory: 1,000 candles, 31 calls and 6 fills

run it yourself:

1 → pip install lightgbm google-adk pytest

2 → one call a day at the 17:00 utc close, logged to prometheus

3 → the boring part: both tests must pass. the calls have to beat holding eth after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have fired the brain in week one. 25 sell calls in 31 days looked broken to me.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

https://t.co/N1Y2Ue0Kzr

[RetroChainer @RetroChainer]

TWO SIGMA PAYS A QUANT RESEARCHER A MEDIAN $350K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

gluonts, camel, promptfoo and influxdb.

82k stars between them, and the eyes are amazon's

for $0 you get:

→ gluonts: probabilistic time series modeling in python

→ camel: a framework for teams of ai agents

→ promptfoo: a test runner for prompts, agents and rag pipelines

→ influxdb: a scalable datastore for metrics, events and real-time analytics

how they fit:

1. gluonts is the eyes: the close landed inside its 90% band on 27 of 30 days

2. camel is the brain: 5 votes, bull vs bear

3. promptfoo is the judge: 30 calls vs just holding btc, after costs

4. influxdb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install gluonts camel-ai

2 → one call a day at the 17:00 utc close, logged to influxdb

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let the brain trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

TWO SIGMA PAYS A QUANT RESEARCHER A MEDIAN $350K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

gluonts, camel, promptfoo and influxdb.

82k stars between them, and the eyes are amazon's

for $0 you get:

→ gluonts: probabilistic time series modeling in python

→ camel: a framework for teams of ai agents

→ promptfoo: a test runner for prompts, agents and rag pipelines

→ influxdb: a scalable datastore for metrics, events and real-time analytics

how they fit:

1. gluonts is the eyes: the close landed inside its 90% band on 27 of 30 days

2. camel is the brain: 5 votes, bull vs bear

3. promptfoo is the judge: 30 calls vs just holding btc, after costs

4. influxdb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install gluonts camel-ai

2 → one call a day at the 17:00 utc close, logged to influxdb

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let the brain trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

HUDSON RIVER TRADING PAYS A QUANT RESEARCHER A MEDIAN $354K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

sktime, qwen3, deepeval and timescaledb.

81k stars between them, and the brain is alibaba's

for $0 you get:

→ sktime: a unified framework for machine learning with time series

→ qwen3: the large language model series from alibaba cloud's qwen team

→ deepeval: the llm evaluation framework

→ timescaledb: a time-series database packaged as a postgres extension

how they fit:

1. sktime is the eyes: the close landed inside its 90% band on 27 of 30 days

2. qwen3 is the brain: 5 votes, bull vs bear

3. deepeval is the judge: 30 calls vs just holding btc, after costs

4. timescaledb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install sktime deepeval

2 → one call a day at the 17:00 utc close, logged to a timescaledb hypertable

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let qwen3 trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

HUDSON RIVER TRADING PAYS A QUANT RESEARCHER A MEDIAN $354K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

sktime, qwen3, deepeval and timescaledb.

81k stars between them, and the brain is alibaba's

for $0 you get:

→ sktime: a unified framework for machine learning with time series

→ qwen3: the large language model series from alibaba cloud's qwen team

→ deepeval: the llm evaluation framework

→ timescaledb: a time-series database packaged as a postgres extension

how they fit:

1. sktime is the eyes: the close landed inside its 90% band on 27 of 30 days

2. qwen3 is the brain: 5 votes, bull vs bear

3. deepeval is the judge: 30 calls vs just holding btc, after costs

4. timescaledb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install sktime deepeval

2 → one call a day at the 17:00 utc close, logged to a timescaledb hypertable

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let qwen3 trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days, and by how much, is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

JANE STREET PAYS A QUANT RESEARCHER A MEDIAN $300K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

statsforecast, deepseek-r1, evals and clickhouse. 168k stars between them. evals is openai's own.

for $0 you get:

→ statsforecast: fast forecasting with statistical and econometric models

→ deepseek-r1: open reasoning models, comparable to openai-o1 per the readme

→ evals: write your own evals for the cases you care about

→ clickhouse: a column-oriented database for real-time analytics

how they fit:

1. statsforecast is the eyes: the close landed inside its 90% band on 27 of 30 days

2. deepseek-r1 is the brain: 5 votes, bull vs bear

3. evals is the judge: 30 calls vs just holding btc, after costs

4. clickhouse is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install statsforecast clickhouse-connect

2 → one call a day at the 17:00 utc close, logged to clickhouse

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let r1 trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

JANE STREET PAYS A QUANT RESEARCHER A MEDIAN $300K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

statsforecast, deepseek-r1, evals and clickhouse. 168k stars between them. evals is openai's own.

for $0 you get:

→ statsforecast: fast forecasting with statistical and econometric models

→ deepseek-r1: open reasoning models, comparable to openai-o1 per the readme

→ evals: write your own evals for the cases you care about

→ clickhouse: a column-oriented database for real-time analytics

how they fit:

1. statsforecast is the eyes: the close landed inside its 90% band on 27 of 30 days

2. deepseek-r1 is the brain: 5 votes, bull vs bear

3. evals is the judge: 30 calls vs just holding btc, after costs

4. clickhouse is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install statsforecast clickhouse-connect

2 → one call a day at the 17:00 utc close, logged to clickhouse

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let r1 trade on day one. 18 sell calls in 31 days and i'd still be arguing with the judge.

what the judge decided after 30 days is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

THE EYES ON THIS BTC DESK PUT 25 OF THE LAST 30 CLOSES INSIDE THEIR RANGE, AND THE JUDGE STILL WON'T RELEASE $1 OF REAL MONEY. ONE OF ITS 4 FREE REPOS IS OPENAI'S.

autogluon, gpt-oss, mlflow and redis. 137k stars between them.

for $0 you get:

→ autogluon: fast and accurate ml in 3 lines of code

→ gpt-oss: two open-weight language models from openai, 120b and 20b

→ mlflow: an open source platform to debug, evaluate and monitor ai apps

→ redis: a fast cache and data structure server for real-time apps

how they fit:

1. the eyes draw 400 paths for btc's next 24 hours

2. the brain counts 5 stated votes, bull vs bear

3. the judge compares runs after fees and slippage

4. the memory streams every call into redis

run it yourself:

1 → pip install autogluon mlflow redis

2 → paste: "xadd desk:calls * day 11-02 call SELL pos cash"

3 → keep the gate shut until the calls beat holding btc after costs

honestly the memory panel is the part i keep watching. one entry per call, 31 so far. i'd have quietly deleted the bad ones.

what the judge decided on day 30 is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

THE EYES ON THIS BTC DESK PUT 25 OF THE LAST 30 CLOSES INSIDE THEIR RANGE, AND THE JUDGE STILL WON'T RELEASE $1 OF REAL MONEY. ONE OF ITS 4 FREE REPOS IS OPENAI'S.

autogluon, gpt-oss, mlflow and redis. 137k stars between them.

for $0 you get:

→ autogluon: fast and accurate ml in 3 lines of code

→ gpt-oss: two open-weight language models from openai, 120b and 20b

→ mlflow: an open source platform to debug, evaluate and monitor ai apps

→ redis: a fast cache and data structure server for real-time apps

how they fit:

1. the eyes draw 400 paths for btc's next 24 hours

2. the brain counts 5 stated votes, bull vs bear

3. the judge compares runs after fees and slippage

4. the memory streams every call into redis

run it yourself:

1 → pip install autogluon mlflow redis

2 → paste: "xadd desk:calls * day 11-02 call SELL pos cash"

3 → keep the gate shut until the calls beat holding btc after costs

honestly the memory panel is the part i keep watching. one entry per call, 31 so far. i'd have quietly deleted the bad ones.

what the judge decided on day 30 is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

CITADEL PAYS A QUANT RESEARCHER A MEDIAN $400K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

darts, dspy, hftbacktest and duckdb.

96k stars between them. dspy started at stanford nlp

for $0 you get:

→ darts: user-friendly forecasting and anomaly detection on time series

→ dspy: compositional python code instead of brittle prompts

→ hftbacktest: a backtester that accounts for queue positions and latencies

→ duckdb: an analytical database, no separate server to manage

how they fit:

1. darts is the eyes: the close landed inside its 90% band on 27 of 30 days

2. dspy is the brain: 5 votes, bull vs bear

3. hftbacktest is the judge: 30 calls vs just holding btc, after costs

4. duckdb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install darts dspy hftbacktest duckdb

2 → one call a day at the 17:00 utc close, logged to desk.duckdb

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let the brain trade on day one. the judge exists for people like me.

what the judge decided after 30 days is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

CITADEL PAYS A QUANT RESEARCHER A MEDIAN $400K A YEAR. THIS BTC DESK GIVES 4 FREE REPOS ONE ROLE EACH, AND ONE OF THEM DECIDES IF A REAL DOLLAR GOES IN.

darts, dspy, hftbacktest and duckdb.

96k stars between them. dspy started at stanford nlp

for $0 you get:

→ darts: user-friendly forecasting and anomaly detection on time series

→ dspy: compositional python code instead of brittle prompts

→ hftbacktest: a backtester that accounts for queue positions and latencies

→ duckdb: an analytical database, no separate server to manage

how they fit:

1. darts is the eyes: the close landed inside its 90% band on 27 of 30 days

2. dspy is the brain: 5 votes, bull vs bear

3. hftbacktest is the judge: 30 calls vs just holding btc, after costs

4. duckdb is the memory: 1,000 candles, 31 calls and 10 fills

run it yourself:

1 → pip install darts dspy hftbacktest duckdb

2 → one call a day at the 17:00 utc close, logged to desk.duckdb

3 → the boring part: the calls must beat holding btc after costs, overall and in 2 of 3 ten-day blocks

honestly i'd have let the brain trade on day one. the judge exists for people like me.

what the judge decided after 30 days is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

2 FREE REPOS, ONE OF THEM GOOGLE'S TENSORFLOW. ONE GROK BOT DESK. $13,584 IN 3 DAYS, AND ONE COIN DID MOST OF THE WORK.

tensorflow and polars. 242k stars between them.

for $0 you get:

→ tensorflow: google's open source machine learning framework

→ polars: an extremely fast dataframe query engine written in rust

→ a live feed of trades and taker buys for pengu, wif and bonk

how they fit:

1. polars rolls the last 48 bars into 6 features

2. the analyst scores long and short from 0 to 100

3. the trader takes one position at a time

run it yourself:

1 → pip install tensorflow polars

2 → load 5-minute bars into a polars dataframe

3 → check the taker-buy share before you trust a score

honestly the coin rows are the part i keep watching. 3 coins, 25 trades, and the split between them surprised me. i'd have bet on a different coin.

which coin did most of it is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

2 FREE REPOS, ONE OF THEM GOOGLE'S TENSORFLOW. ONE GROK BOT DESK. $13,584 IN 3 DAYS, AND ONE COIN DID MOST OF THE WORK.

tensorflow and polars. 242k stars between them.

for $0 you get:

→ tensorflow: google's open source machine learning framework

→ polars: an extremely fast dataframe query engine written in rust

→ a live feed of trades and taker buys for pengu, wif and bonk

how they fit:

1. polars rolls the last 48 bars into 6 features

2. the analyst scores long and short from 0 to 100

3. the trader takes one position at a time

run it yourself:

1 → pip install tensorflow polars

2 → load 5-minute bars into a polars dataframe

3 → check the taker-buy share before you trust a score

honestly the coin rows are the part i keep watching. 3 coins, 25 trades, and the split between them surprised me. i'd have bet on a different coin.

which coin did most of it is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

https://t.co/N1Y2Ue0Kzr

[RetroChainer @RetroChainer]

ONE WIF SHORT PAID $4,594. THE GROK BOT DESK BEHIND IT IS BUILT ON 2 FREE REPOS, ONE OF THEM HUGGING FACE'S, AND IT SAT IN THAT TRADE FOR HOURS.

transformers and pandas. 219k stars between them.

for $0 you get:

→ transformers: hugging face's framework for machine learning models in text, vision and audio

→ pandas: labeled data structures and statistical functions for python

→ trades and volume for pengu, wif and bonk on every 5-minute bar

how they fit:

1. the feed counts every trade per bar. 268 pengu trades in the last one

2. 6 features score the last 4 hours, and nothing trades under 70

3. a trail closes the winners. 16 trail exits so far

run it yourself:

1 → pip install transformers pandas

2 → resample the trades into 5-minute bars with pandas

3 → watch the trail line before you trust a trade

honestly the orders table is the part i keep watching. 25 trades, 17 green. i'd have closed that wif short at half the move.

exactly how many bars it held that short is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

ONE WIF SHORT PAID $4,594. THE GROK BOT DESK BEHIND IT IS BUILT ON 2 FREE REPOS, ONE OF THEM HUGGING FACE'S, AND IT SAT IN THAT TRADE FOR HOURS.

transformers and pandas. 219k stars between them.

for $0 you get:

→ transformers: hugging face's framework for machine learning models in text, vision and audio

→ pandas: labeled data structures and statistical functions for python

→ trades and volume for pengu, wif and bonk on every 5-minute bar

how they fit:

1. the feed counts every trade per bar. 268 pengu trades in the last one

2. 6 features score the last 4 hours, and nothing trades under 70

3. a trail closes the winners. 16 trail exits so far

run it yourself:

1 → pip install transformers pandas

2 → resample the trades into 5-minute bars with pandas

3 → watch the trail line before you trust a trade

honestly the orders table is the part i keep watching. 25 trades, 17 green. i'd have closed that wif short at half the move.

exactly how many bars it held that short is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG

[RetroChainer @RetroChainer]

XAI PUT GROK-1 ON GITHUB FOR FREE. THE GROK BOT DESK IN THIS CLIP MADE $4,993 IN 3 DAYS, AND ITS OWN TABLE SHOWS AN EXIT THAT MADE $7,929.

grok-1 and scikit-learn. 120k stars between them.

for $0 you get:

→ grok-1: xai's open release, 314b parameters, apache 2.0

→ scikit-learn: machine learning in python

→ 6 roles on one screen: feed, features, analyst, trader, risk, logger

how they fit:

1. 6 features read the last 4 hours of each coin on every 5-minute bar

2. one position at a time, 8 hours max

3. every fill gets logged, 165 lines so far

run it yourself:

1 → pip install scikit-learn

2 → build the features on the last 48 bars

3 → re-run the same signals with other exits before you trust one

honestly the exits table is the part i keep watching. same signals, 5 other ways out. i'd have stopped tuning after the first one.

which exit it was is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

XAI PUT GROK-1 ON GITHUB FOR FREE. THE GROK BOT DESK IN THIS CLIP MADE $4,993 IN 3 DAYS, AND ITS OWN TABLE SHOWS AN EXIT THAT MADE $7,929.

grok-1 and scikit-learn. 120k stars between them.

for $0 you get:

→ grok-1: xai's open release, 314b parameters, apache 2.0

→ scikit-learn: machine learning in python

→ 6 roles on one screen: feed, features, analyst, trader, risk, logger

how they fit:

1. 6 features read the last 4 hours of each coin on every 5-minute bar

2. one position at a time, 8 hours max

3. every fill gets logged, 165 lines so far

run it yourself:

1 → pip install scikit-learn

2 → build the features on the last 48 bars

3 → re-run the same signals with other exits before you trust one

honestly the exits table is the part i keep watching. same signals, 5 other ways out. i'd have stopped tuning after the first one.

which exit it was is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

https://t.co/N1Y2Ue0Kzr

[RetroChainer @RetroChainer]

A GROK BOT DESK MADE $1,898 IN 3 DAYS BETTING AGAINST MEMECOIN MOVES. IT SCORES EVERY COIN OUT OF 100, AND ONE NUMBER IS WORTH 30 OF THOSE POINTS.

cryptofeed and tsfresh.

12k stars between them

for $0 you get:

→ cryptofeed: a websocket data feed handler for crypto exchanges

→ tsfresh: automatic extraction of relevant features from time series

→ a 0-100 fade score for pengu, wif and bonk on every 5-minute bar

how they fit:

1. the feed counts trades and volume per bar. 360 trades in the last one

2. 6 features score the last 4 hours. 17,118 values so far

3. nothing trades under 70

run it yourself:

1 → pip install cryptofeed tsfresh

2 → stream binance trades for 3 coins into 5-minute bars

3 → score the last 48 bars before you trust a signal

honestly the features table is the part i keep watching. 6 numbers decide every trade. i'd have gone with my gut on half of them.

which number carries 30 of the 100 points is on screen in the clip (visualization).

bookmark this ↓

[RetroChainer @RetroChainer]

A GROK BOT DESK MADE $1,898 IN 3 DAYS BETTING AGAINST MEMECOIN MOVES. IT SCORES EVERY COIN OUT OF 100, AND ONE NUMBER IS WORTH 30 OF THOSE POINTS.

cryptofeed and tsfresh.

12k stars between them

for $0 you get:

→ cryptofeed: a websocket data feed handler for crypto exchanges

→ tsfresh: automatic extraction of relevant features from time series

→ a 0-100 fade score for pengu, wif and bonk on every 5-minute bar

how they fit:

1. the feed counts trades and volume per bar. 360 trades in the last one

2. 6 features score the last 4 hours. 17,118 values so far

3. nothing trades under 70

run it yourself:

1 → pip install cryptofeed tsfresh

2 → stream binance trades for 3 coins into 5-minute bars

3 → score the last 48 bars before you trust a signal

honestly the features table is the part i keep watching. 6 numbers decide every trade. i'd have gone with my gut on half of them.

which number carries 30 of the 100 points is on screen in the clip (visualization).

bookmark this ↓

[Сarm1ne @carm1nee]

these 10 GitHub repos are the most DANGEROUS thing in trading right now

together they give you a full AI hedge fund: a research desk, an investment committee and an execution desk

in this article I've shown EXACTLY how to build it from scratch https://t.co/17pf1e63VG