A research-first crypto trading framework for BTC/USDT and other liquid crypto markets, built around one simple idea: a beautiful backtest is not a real edge.
FlowMate is designed to test whether a strategy can survive real costs, real market structure, realistic execution, and statistical validation before any serious capital is risked.
(Magyar verzió lentebb — ugrás a magyar leíráshoz)
FlowMate is an active research and engineering project. It is not a finished production trading system.
The current development focus is:
- realistic orderflow and execution backtesting;
- causal orderflow and indicator primitives;
- strategy registry and strategy-search infrastructure;
- validation gates against overfitting;
- fee-aware and turnover-aware objective functions;
- thermal-safe and performance-aware matrix runs;
- research-only strategy candidates;
- future DEX and arbitrage research through adapters, not a separate bot.
Live/paper defaults must remain disabled until the relevant validation, replay, execution and review gates pass.
FlowMate is a Python-based crypto strategy research and trading-bot framework.
Its near-term purpose is to help answer:
Does this strategy still look useful after fees, slippage, funding, partial fills,
latency, bad regimes, data-quality degradation and overfitting checks?
FlowMate can support signal/paper/live workflows, but the project policy is conservative:
research -> realistic backtest -> validation -> paper observation -> small controlled live pilot
No strategy should be promoted just because it has a strong in-sample backtest.
FlowMate is not:
- a guaranteed-profit system;
- a “daily average profit” promise;
- a finished live trading product;
- a shortcut around research and risk management;
- a financial-advice tool.
Backtested performance is treated as suspicious until it survives realistic execution and out-of-sample validation.
Most trading bots show a good-looking equity curve and a large number. FlowMate starts from the opposite assumption:
That number is probably wrong until proven otherwise.
Three principles guide the project:
- Research first. A strategy is a hypothesis, not a promise.
- Risk first. Capital protection matters more than chasing headline returns.
- Realistic testing first. Fees, slippage, funding, partial fills, latency, data quality and overfitting must be modeled before a strategy is trusted.
A strategy candidate should pass a strict acceptance gate before it moves beyond research status.
FlowMate separates strategy logic into layers:
HTF bias / regime 4h–1D trend, momentum, ADX/HMM regime
MTF setup candidate 15m–1h breakout, pullback, squeeze, value area
LTF orderflow gate 1m–5m OFI, CVD, absorption, sweep, microprice
Risk layer all TFs ATR stop, vol targeting, funding/VPIN veto
Validation layer offline matrix, CPCV, DSR, PBO, realistic execution
The intended design is:
Indicators find where and in which direction to look.
Orderflow decides whether entry timing is acceptable now.
Risk and validation decide whether the idea is worth keeping.
Orderflow is not treated as a standalone signal factory by default. It is primarily a confirmation, timing, filter or veto layer for larger target setups.
A strategy can move from research to validated only if it passes conditions such as:
- causal logic, no
shift(-1), no future rows and no full-history normalization; - explicit data-quality requirements and
data_quality_scorehandling; - realistic execution costs: fees, slippage, funding, spread, partial fills, rejected fills and latency;
- matrix robustness: worst-case cell, consistency, live-proxy score and scenario stress;
- walk-forward or CPCV validation with purging and embargo;
- Deflated Sharpe / PBO checks;
- sufficient trade count and sample size;
- regime robustness and parameter stability;
- structured signal reasons:
entry_reason,exit_reason,invalidation_reason,risk_veto_reason.
Even then:
validated != live default
validated != paper default
Validation means a research gate was passed, not that a strategy is ready for uncontrolled deployment.
There is no reliable “daily average profit” number.
A strategy that is useful in research may still fail in live conditions because of:
- hidden future information in a backtest;
- optimistic fill assumptions;
- too many parameter trials;
- regime dependency;
- fees, slippage and funding;
- latency and missed fills;
- live data quality issues.
The right question is not:
How much does it make per day?
The right question is:
Does it remain positive across regimes, after all costs, without relying on one lucky parameter set?
The only honest profit estimate comes from a carefully controlled pilot with real execution, and even that is a range, not a fixed number.
The current roadmap is organized through console handoff files under:
docs/ai_handoff/
Start here:
git fetch origin
git checkout main
git pull --ff-only
cat docs/old/ai_handoff_archive/CONSOLE_START_HERE_ALL_PROJECTS_2026_05_30.mdActive tracks include:
- Orderflow deepening + clean data ingestion
- causal primitives, EventTape, local book reconstruction, feed reliability.
- Orderflow research addendum
- MLOFI, microprice, VPIN, diagonal/stacked imbalance, FDR/iceberg proxy.
- Strategy search + optimizer + strategy zoo
- registry, CPCV/DSR/PBO, Optuna over robust objectives, research-only candidates.
- Backtest Thermal Guard
- safe CPU/temperature throttling for heavy matrix runs.
- Backtest performance optimization
- profile first, optimize only measured hot paths.
- Polars + Numba staged migration
- optional acceleration with fallback and equivalence tests.
- v5.7 realistic orderflow backtest
- non-trivial orderflow features, realistic execution and legacy-vs-realistic comparison.
- Indicator expansion + fee-aware confluence
- HTF momentum/breakout indicators, turnover control, fee-aware objective functions.
- DEX + arbitrage foundation
- future adapter/fill-engine/scenario work for Solana, Sui and Base; arbitrage later.
The current 30-day MVP plan is not to build a finished live trading system.
The target is:
A working strategy research + validation engine with tested orderflow primitives,
thermal-safe research runs, a research-only strategy registry, first validation gates,
and the first realistic execution/replay foundations.
Read:
cat docs/old/ai_handoff_archive/ONE_MONTH_MVP_EXECUTION_PLAN_2026_05_30.mdRecommended order:
1. Orderflow primitive foundation.
2. Orderflow primitive tests.
3. Indicator/fee confluence primitives and turnover controls.
4. Strategy registry skeleton.
5. Validation skeleton.
6. Thermal Guard before heavy matrix/search runs.
7. Profiling before performance optimization.
8. Fee-aware objective and optimizer skeleton only after validation exists.
9. Live data ingestion only after primitive/replay safety is in place.
10. DEX primitive foundation only after the core CEX research stack is stable.
11. Arbitrage only after DEX fill engines and MEV/order-race scenarios exist.
Use a virtual environment:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txtRun compile/test checks:
python -m compileall -q crypto_bot tests tools research
pytest -qFor heavy matrix or optimization work, use conservative environment settings:
FLOWMATE_MAX_WORKERS=1
FLOWMATE_MAX_CPU_PCT=80
FLOWMATE_MAX_TEMP_C=82
POLARS_MAX_THREADS=2
OMP_NUM_THREADS=1
MKL_NUM_THREADS=1
OPENBLAS_NUM_THREADS=1
NUMEXPR_NUM_THREADS=1Do not commit generated data/results:
Optuna DBs
matrix CSVs
Parquet data
profiles
ZIPs
.env files
virtualenvs
local caches
For ChatGPT Codex console and Claude console:
git fetch origin
git checkout main
git pull --ff-only
cat docs/old/ai_handoff_archive/CONSOLE_START_HERE_ALL_PROJECTS_2026_05_30.mdThen choose exactly one project branch and one PR-sized task.
Rules:
one branch
one task
one review
one merge
Claude review should check:
- no lookahead;
- no live/paper activation;
- sibling-only strategy additions;
- correct aggressor signs and data-quality rules;
- overfitting controls;
- realistic cost handling;
- tests and rollback path.
FlowMate’s current CEX work is Binance-focused, with spot and futures research paths.
Future DEX work is planned as:
DEX = adapter + fill-engine + scenario extension
It is not a separate bot, and early DEX work must be read-only / fixture-based:
No wallet signing.
No private keys.
No live RPC default.
No arbitrage execution in early PRs.
Arbitrage is a later research subsystem and must model gas, priority fees, slippage, revert costs, order-race/MEV competition, and inventory/bridge risk before any result is trusted.
The project is designed around conservative assumptions:
- small controlled pilots only after validation;
- low leverage if futures are ever used;
- explicit stop/risk logic;
- daily loss limits;
- net performance after costs;
- false negatives are preferred over optimistic fills.
Leverage amplifies losses as much as gains. Use the smallest possible exposure while testing.
This is a research and engineering project.
It is not financial advice. It is not a guarantee of profit. Trading cryptocurrencies, especially with leverage, can lose money quickly. Past performance and backtested performance do not predict future results.
Only trade with capital you can afford to lose, and start with small controlled pilots only after the research and validation gates are complete.
Kutatás-központú kripto trading framework BTC/USDT-re és más likvid kripto piacokra, egy egyszerű alapelvvel: a szép backtest még nem valódi edge.
A FlowMate célja annak vizsgálata, hogy egy stratégia túléli-e a valós költségeket, a valós piaci mikrostruktúrát, a realisztikus teljesülést és a statisztikai validációt, mielőtt komoly tőke kerülne kockára.
A FlowMate aktív kutatási és mérnöki projekt. Nem kész production live trading rendszer.
A jelenlegi fejlesztési fókusz:
- realisztikus orderflow és execution backtest;
- kauzális orderflow- és indikátor-primitívek;
- strategy registry és stratégia-kereső infrastruktúra;
- overfitting elleni validációs kapuk;
- fee- és turnover-tudatos objektív függvények;
- thermal-safe és performance-aware matrix futások;
- research-only stratégiajelöltek;
- későbbi DEX és arbitrázs kutatás adaptereken keresztül.
Live/paper default nem változhat addig, amíg a megfelelő validációs, replay, execution és review kapuk nem teljesülnek.
A FlowMate Python-alapú kripto stratégia-kutatási és trading-bot framework.
A rövid távú célja erre válaszolni:
Használható marad-e ez a stratégia díjak, slippage, funding, részleges teljesülés,
latency, rossz rezsim, adatminőség-romlás és overfitting-szűrés után is?
A rendszer támogathat signal/paper/live workflow-kat, de a projektpolitika konzervatív:
research -> realisztikus backtest -> validáció -> paper megfigyelés -> kis kontrollált live pilot
Semmi nem kerülhet tovább csak azért, mert szép in-sample backtestet mutat.
A FlowMate nem:
- garantált profit rendszer;
- napi átlag profit ígéret;
- kész live trading termék;
- kutatást és kockázatkezelést kiváltó shortcut;
- pénzügyi tanácsadó eszköz.
A backtest eredmény gyanús, amíg nem él túl realisztikus executiont és out-of-sample validációt.
A legtöbb trading bot szép equity curve-öt és nagy számot mutat. A FlowMate az ellenkező feltételezésből indul ki:
Ez a szám valószínűleg téves, amíg be nem bizonyítjuk az ellenkezőjét.
Három alapelv:
- Előbb a kutatás. Egy stratégia hipotézis, nem ígéret.
- Előbb a kockázat. A tőke védelme fontosabb, mint a headline hozam.
- Előbb a realisztikus teszt. Fee, slippage, funding, részleges teljesülés, latency, adatminőség és overfitting nélkül nem bízunk a számban.
HTF bias / rezsim 4h–1D trend, momentum, ADX/HMM rezsim
MTF setup candidate 15m–1h breakout, pullback, squeeze, value area
LTF orderflow gate 1m–5m OFI, CVD, absorption, sweep, microprice
Risk layer minden TF ATR stop, vol targeting, funding/VPIN veto
Validation layer offline matrix, CPCV, DSR, PBO, realistic execution
Az indikátor megmondja, hol és merre érdemes nézni. Az orderflow azt mondja meg, hogy érdemes-e most belépni. A risk és validation réteg dönti el, hogy az ötletet meg szabad-e tartani.
Egy stratégia csak akkor léphet research státuszból validated státuszba, ha többek között teljesül:
- kauzális logika, nincs
shift(-1), nincs jövőbeli sor, nincs full-history normalizálás; - explicit data-quality követelmény és
data_quality_scorekezelés; - realisztikus execution: fee, slippage, funding, spread, partial/rejected fills, latency;
- matrix robustness: worst-case, consistency, live-proxy score, scenario stress;
- walk-forward vagy CPCV validáció purging/embargo mellett;
- Deflated Sharpe / PBO ellenőrzés;
- elég trade és elég minta;
- rezsim-robusztusság és paraméter-stabilitás;
- strukturált okmezők:
entry_reason,exit_reason,invalidation_reason,risk_veto_reason.
Fontos:
validated != live default
validated != paper default
Nincs megbízható „napi átlag profit” szám.
Egy stratégia researchben jól nézhet ki, de éles környezetben elbukhat:
- rejtett future leak miatt;
- optimista fill feltételezések miatt;
- túl sok paraméterpróba miatt;
- rezsimfüggőség miatt;
- fee, slippage, funding miatt;
- latency és missed fill miatt;
- live adatminőség miatt.
A helyes kérdés nem az, hogy:
Mennyit csinál naponta?
Hanem:
Pozitív marad-e több rezsimen át, minden költség után, anélkül,
hogy egy szerencsés paraméterre támaszkodna?
A jelenlegi roadmap a docs/ai_handoff/ alatt található.
Indulás:
git fetch origin
git checkout main
git pull --ff-only
cat docs/old/ai_handoff_archive/CONSOLE_START_HERE_ALL_PROJECTS_2026_05_30.mdAktív trackek:
- Orderflow deepening + clean data ingestion.
- Orderflow research addendum.
- Strategy search + optimizer + strategy zoo.
- Backtest Thermal Guard.
- Backtest performance optimization.
- Polars + Numba staged migration.
- v5.7 realistic orderflow backtest.
- Indicator expansion + fee-aware confluence.
- DEX + arbitrage foundation.
A jelenlegi 30 napos cél nem kész live trading rendszer.
A cél:
Működő strategy research + validation engine, tesztelt orderflow primitivekkel,
thermal-safe kutatási futásokkal, research-only registryvel, első validációs kapukkal,
és az első realistic execution/replay alapokkal.
Olvasd:
cat docs/old/ai_handoff_archive/ONE_MONTH_MVP_EXECUTION_PLAN_2026_05_30.mdpython3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -r requirements.txt
python -m compileall -q crypto_bot tests tools research
pytest -qHeavy matrix / optimizer futások előtt:
FLOWMATE_MAX_WORKERS=1
FLOWMATE_MAX_CPU_PCT=80
FLOWMATE_MAX_TEMP_C=82
POLARS_MAX_THREADS=2
OMP_NUM_THREADS=1
MKL_NUM_THREADS=1
OPENBLAS_NUM_THREADS=1
NUMEXPR_NUM_THREADS=1ChatGPT Codex console és Claude console mindig ezzel induljon:
git fetch origin
git checkout main
git pull --ff-only
cat docs/old/ai_handoff_archive/CONSOLE_START_HERE_ALL_PROJECTS_2026_05_30.mdSzabály:
egy branch
egy task
egy review
egy merge
Ez kutatási és mérnöki projekt.
Nem pénzügyi tanácsadás. Nem profitgarancia. A kriptokereskedés, különösen tőkeáttéttel, gyorsan veszteséget okozhat. A múltbeli és backtestelt teljesítmény nem jelzi előre a jövőbeli eredményeket.
Csak olyan tőkével kereskedj, amelynek elvesztését megengedheted magadnak, és csak a kutatási/validációs kapuk után kezdj kis, kontrollált pilottal.