Prediction under uncertainty, tested against reality.
Synapse is a research and prediction harness for models operating in messy market conditions. Powered by live context, structured disagreement, and a logged thought process that can be inspected after the outcome is known, to enable you to improve your principles.
Trading is the first testbed. The bigger goal is general prediction under uncertainty.
You · 14:02
Build a low-vol momentum signal on the S&P 500. Use 12-1 momentum with volatility scaling, rebalance monthly, and report Sharpe & drawdown since 2005.
planner → outline ready · 6 cells · data: yfinance, FRED
coder ✓ cell 1/6 · load universe
coder ✓ cell 2/6 · compute 12-1 momentum
coder ✓ cell 3/6 · volatility scaling
tester ▶ running cell 4/6
Sharpe
1.42
Max DD
−14.8%
Turnover
87%
Harness lenses
The workflow
The harness matters as much as the model.
The edge is not just which model answers. It is what the system forces the model to see, what it prevents it from seeing, how it debates uncertainty, and how every claim is turned into something measurable.
- 01
State the uncertainty
Start with a market hypothesis, forecast question, or repeatable strategy idea instead of a benchmark-style prompt.
- 02
Force the lenses
Synapse routes the model through data, news, fundamentals, technical context, recipes, and debate so confidence has to earn its way through constraints.
- 03
Log the artifact
Every notebook, run, forecast, debate, signal, and paper trade lands in the workspace so the reasoning can be audited against reality later.
Inside a session
Not a market take. A testable object.
The goal is not a polished paragraph that sounds plausible. The goal is a notebook, recipe run, debate record, forecast, or signal view that can be opened, inspected, rerun, and scored when reality catches up.
- Notebook artifacts. Generated notebooks include markdown, code, provenance metadata, validation events, lineage extraction, and bias checks.
- Recipe artifacts. YAML recipes compile into notebooks, execute through a Jupyter kernel, and persist metrics, folds, regimes, tags, and searches.
- Decision artifacts. Debates and forecasts keep the analyst lenses, evidence, conviction, resolution criteria, and performance view together.
## Low-vol momentum on the S&P 500 Build a long-only, volatility-scaled 12-1 momentum signal. Rebalance monthly. Evaluate 2005-present.
from findata.equity_prices import get_equity_prices from findata.index_constituents import get_sp500_composition universe = get_sp500_composition(asof="2005-01-01") prices = get_equity_prices(universe.symbol, start_date="2005-01-01")
returns = prices["close"].pct_change() mom_12_1 = prices["close"].pct_change(252).shift(21) vol = returns.rolling(63).std() signal = (mom_12_1 / vol).rank(axis=1, pct=True)
### Portfolio construction
weights = (signal > 0.8).astype(float) weights = weights.div(weights.sum(axis=1), axis=0).fillna(0) strategy_returns = (weights.shift(1) * returns).sum(axis=1)
Security
Built so teams can audit the process, not just the answer.
Synapse is in private beta, with an emphasis on readable boundaries, inspectable artifacts, and deployment choices that can harden with each research team.
Inspectable boundaries
The frontend, orchestrator, data MCP, code MCP, and live-data server communicate over explicit HTTP, MCP, and WebSocket boundaries.
Models are swappable
Synapse is being built so teams can choose the model and provider that fit each workflow while the harness and artifacts stay stable.
Self-hostable stack
Synapse is designed for private deployments with clear service ownership, health checks, internal networking, and observability hooks.
Claims leave a trail
Runs, searches, debates, costs, lineage, signal series, forecasts, and notebook outputs are stored so research can be inspected after generation.
FAQ
Straight answers.
What is Synapse today?+
Synapse is a research and prediction harness for testing model reasoning under uncertainty. Today it includes chat-driven notebook generation, deterministic recipe runs, debate records, event forecasts, live-data dashboards, signal views, paper trading, and artifact history.
Is Synapse model agnostic?+
Yes. Synapse is built as a model-agnostic harness: the constraints, artifacts, and review loop stay consistent regardless of what runs underneath. Custom models may be made available for on-prem or customised deployments.
Where does the data come from?+
Equity prices come from yfinance and broker APIs, macro series from FRED, and news and sentiment from GDELT. Research workflows reach them through findata and data-mcp — documented wrappers across equity prices, macro, factors, crypto, volatility, files, and diagnostics — while live streams add market status, corporate events, holdings, and governance context across supported markets.
Can I self-host it?+
Yes. Synapse is designed for private deployment, with the web app and backend services kept behind clear internal service boundaries. Reach the founders at founders@synapsetrading.net.
How do I get access?+
Synapse is in private beta. Request early access and we will onboard research teams whose workflows match the current product. Reach the founders at founders@synapsetrading.net.
Help us test the harness.
Request access if you are building systematic strategies, forecast workflows, or model-evaluation loops where evidence, constraints, and logged outcomes matter more than a confident paragraph.