Off-the-shelf — deploy anywhere in hours

Your competitors are trading on microstructure data you don't have.

Tagashira is a ready-to-deploy market data platform for crypto derivatives. 69 labelled, ML-ready features. Sub-200ms latency. Ships with Jupyter notebooks for signal generation, model training, and walk-forward validation. Runs on a $40/month server. Live in hours, not months.

Request Technical Brief → See the Code

The Problem

The average quant desk spends
6-9 months building data infra
before writing a single signal.

That is senior engineering talent burning runway on plumbing, not alpha. Tagashira is the off-the-shelf alternative.

Silent WebSocket Failures

Connections drop during high-volatility windows, exactly when data matters most. Without robust reconnection logic, strategies operate blind at critical moments.

Rate Limit Exhaustion

Exchange REST APIs impose aggressive quotas. Naive polling burns through limits in minutes, blocking access to funding rates, open interest, and contract specifications.

No ML-Ready Features

Commercial feeds deliver raw OHLCV. Building 69 derived features requires months of feature engineering. Training ML models requires labelled, structured data that raw feeds do not provide.

No Research Environment

Even with data, teams need Jupyter notebooks, signal libraries, backtesting frameworks, and walk-forward validation. Building this research stack from scratch doubles the timeline.

Tagashira eliminates all of this.

A single Docker deployment replaces months of custom engineering. Off the shelf. Running on commodity hardware. Your quant team opens a Jupyter notebook on day one, not a GitHub repo of half-built connectors.

Data arrives labelled and structured. Features are computed statelessly. Parquet output plugs directly into Pandas, Polars, XGBoost, PyTorch, or any ML pipeline.

Hours

To Full Deploy

4 Core

Min. Server Spec

<12%

API Quota Used

~$40

Monthly Infra

Infrastructure Economics

Institutional-grade collection
on a coffee-budget server.

No proprietary hardware. No vendor lock-in. Deploy on any cloud, VPS, NAS, or bare-metal box anywhere in the world. Total cost of ownership: under $500/year.

4 cores

Minimum CPU

Handles 50+ instruments with headroom. 2 cores sufficient for smaller sets.

8 GB

Recommended RAM

4 GB works for fewer than 10 instruments. No memory-hungry dependencies.

~1.5 GB

Daily Storage

Parquet + Zstandard. 5-10x smaller than raw JSON. 100 GB lasts months.

$30-50

Monthly VPS Cost

Hetzner, DigitalOcean, Contabo, or on-premises. Your choice entirely.

Deploy Anywhere in Hours

Ships as a Docker Compose stack. Pull the image, set your API keys, run docker compose up -d. Data collection begins immediately. The monitoring dashboard is live at port 8080. No configuration wizard. No installation scripts. It just works.

Scale Without Rearchitecting

Add instruments at runtime via the dashboard. No restart, no downtime. Need more coverage? Add a second node. Need a second exchange? Drop in a connector module. The architecture scales horizontally without touching a line of core code.

Customise Everything

Every feature, every polling interval, every storage path is configurable. Enterprise clients get full source code to build proprietary signal modules directly into the pipeline. Your data stays on your infrastructure, always.

Who This Is For

Built for professionals who
measure edge in basis points.

Tagashira is not a retail product. It is infrastructure for teams that take data, and the signals they derive from it, seriously.

Systematic Trading Desks

Build and backtest quantitative strategies with full microstructure context. Classical signals (SMA, momentum, mean-reversion) and ML-driven signals (XGBoost, LightGBM ensemble with confidence scoring) from the same labelled feature set. Walk-forward validation prevents overfitting.

Signal generation from labelled features

Market Makers

Order book gradient analysis, depth-weighted mid pricing, and real-time anomaly detection provide continuous market state awareness. Composite pressure scores and funding regime classification inform dynamic spread and inventory management in production.

Sub-200ms data freshness

ML and Quantitative Research

69 labelled features per instrument per 30-second snapshot, output as compressed Parquet. Plug directly into Jupyter, scikit-learn, XGBoost, LightGBM, PyTorch, or any training pipeline. Stateless computation guarantees zero look-ahead bias for backtesting and model validation.

Training-ready feature vectors

Platform Capabilities

From raw WebSocket to
training-ready vectors. Automated.

Everything between the exchange API and your Jupyter notebook, handled. Data arrives labelled, structured, and compressed.

Real-Time Streaming

Persistent WebSocket connections with automatic reconnection and exponential backoff. Full L2 order book, individual trades, and mark/index pricing. Sub-200ms end-to-end latency.

Microstructure Enrichment

Configurable REST poller capturing funding rates, open interest, contract specs, spot reference, deep book snapshots. 30-second cycles consuming less than 12% of exchange API quota.

Order Book Analytics

Multi-level depth analysis with imbalance ratios, slope asymmetry, wall detection, depth-weighted mid pricing, and absorption rate computation across configurable horizons.

Trade Flow Decomposition

Buyer/seller volume decomposition, cumulative volume delta, VWAP deviation, and trade size distribution. Leading indicators for directional and mean-reversion strategies.

Funding and Basis Intelligence

Discrete funding regime classification, crowding detection, carry trade analysis, and spot-futures basis in absolute, bps, and annualised representations. Critical for systematic carry and relative-value strategies.

ML Training Pipeline

All 69 features are labelled, normalised, and output as compressed Parquet. Native input for XGBoost, LightGBM, PyTorch, and scikit-learn. Ships with Jupyter notebooks, signal generators, vectorised backtester, and walk-forward validation framework.

Integration

From deploy to signal
in minutes, not months.

The entire research stack ships with the platform. Load Parquet data, generate signals, train models, and backtest, all from a single Jupyter notebook.

load_features.py

import pandas as pd
from pathlib import Path
# Load labelled microstructure snapshots
df = pd.read_parquet(
    "data/features/BTC-USDT/",
    columns=[
        "timestamp", "close",
        "funding_rate", "l1_imbalance",
        "cvd_ratio", "basis_ann_pct",
        "bid_depth_gradient",
        "composite_pressure_score",
    ]
)
print(f"{len(df):,} rows x {df.shape[1]} features")
# Build training targets
X = df.drop(["timestamp","close"], axis=1)
y = (df.close.shift(-12) / df.close - 1)
y_label = (y > 0).astype(int)

signal_lab.ipynb

from tagashira import DataPool, Signals
from tagashira import Backtester
import xgboost as xgb
# Load 69 features + OHLCV
pool = DataPool()
df = pool.load_merged("BTC-USDT", days=30)
# Traditional signal
sig = Signals.mean_reversion(df, z=2.15)
# ML signal from labelled features
model = xgb.XGBClassifier()
model.fit(X_train, y_train)
ml_sig = model.predict(X_test)
# Backtest with transaction costs
bt = Backtester(cost_bps=5)
bt.run(df, ml_sig).report()
# Sharpe: 1.84 | MaxDD: -4.2%

Signal Generators

SMA, EMA, RSI, momentum,
mean-reversion, breakout, OB imbalance

XGB+LGB

ML Ensemble

Directional prediction with
confidence scoring and time-series CV

DQN

RL Agent

Deep Q-Network with
drawdown-penalised reward

WFV

Walk-Forward

Rolling IS/OOS validation
to prevent overfitting

Architecture

Four containers.
One deploy command.

Docker Compose stack. Pull, configure, run. Live in hours on any Linux host: cloud, VPS, NAS, or bare-metal.

WS Collector

Real-time streams

REST Poller

Microstructure enrichment

Data Aggregator

Parquet + Zstandard

Monitoring Dashboard

Zero-config ops console

Off the Shelf, Not Off the Rack

Ships as a complete product, not a framework, not a boilerplate. Works out of the box with zero custom code. But every component is configurable, and Enterprise clients get full source access to extend it however they need.

Exchange-Agnostic Connectors

Modular connector architecture handles authentication, rate limiting, and message parsing. Adding a new exchange requires a single connector module, no core changes.

Hot-Reload Instruments

Add or remove instruments at runtime through the dashboard. Zero downtime. Scale from 5 to 50+ instruments on a single instance without restarting any container.

Solutions

Three tiers. Each one
solves a different problem.

Perpetual licence with annual support. Every tier includes the full data pipeline, Jupyter research environment, signal library, ML training tools, and backtesting framework.

Standard

Research and Exploration

For quant teams exploring microstructure alpha on a single exchange. Full pipeline, Jupyter notebooks, ML training environment. Deploy in hours.

Up to 10 instruments
Single server deployment
One exchange connector
Full 69-feature pipeline
Jupyter research environment
Signal library + vectorised backtester
ML training tools (XGBoost, LightGBM)
Documentation and email support
Assisted deployment
Source code access

Request Pricing

Stop building plumbing.
Start building alpha.

Request a technical brief or schedule a walkthrough. We respond within 24 hours.

→ Detailed technical brief tailored to your use case
→ Live walkthrough of Jupyter environment and production dashboards
→ Custom deployment plan with infrastructure sizing
→ No-obligation pricing based on your requirements
→ Access to sample Parquet datasets and feature documentation

Request Access

Onboarding limited to 3 new clients per quarter to ensure deployment quality.

Full Name *

Email *

Company

Role

Interested Tier

Timeline

Message (optional)

I confirm I am a human

Verification required to prevent automated submissions

I agree to the Privacy Policy and consent to ETIX storing my information to respond to this enquiry. Your data is never shared with third parties.

Something went wrong. Please email us directly at info@etix.tech

✓

Request Received

Thank you for your interest in Tagashira. Our team will review your submission and respond within 24 hours with a tailored technical brief and deployment plan.