Atlas

Atlas turns global macro noise into explainable action with live signals, scenario simulation, portfolio-aware analysis, and memory for faster financial decisions.

Comment

Atlas

Atlas is a macro decision OS for teams that need to turn fast-moving global events into defensible action. Instead of stopping at summarization, Atlas combines live signal detection, deterministic risk math, scenario simulation, and institutional memory in one workflow. The gitHub repository can be accessed via this link: https://github.com/LordeSid-04/quant-engine.git

Backend Framework

Atlas uses a modular FastAPI backend with a shared API router, typed schemas via Pydantic, and a managed lifespan that starts and stops live market streams cleanly. Data is ingested from multiple market and news providers, persisted through Supabase, and processed through specialized engines:

  • WorldPulseEngine computes the live global factor state and cross-country spillover map.
  • ThemeEngine ingests and classifies news into macro themes.
  • RiskEngine aggregates factor pressure into portfolio-relevant risk categories.
  • ScenarioEngine simulates macro shocks through a directed propagation graph.
  • AnalogueEngine retrieves similar historical regimes.
  • BriefingEngine fuses these outputs into concise, decision-ready narratives.

This architecture is intentionally hybrid: deterministic where trust and auditability matter, and model-assisted where ranking and synthesis add value.

Core Algorithms

1. Cross-Asset Factor Extraction

We normalize market moves from live quotes and map them into latent macro factors such as rates, FX, liquidity, inflation, commodity, and geopolitics.

$$ z_s = \mathrm{clamp}\left(\frac{\frac{p^{close}_s - p^{open}_s}{|p^{open}_s|}}{\sigma_s}, -3, 3\right) $$

$$ f_k = \sum_{s \in S} w_{s,k} \, z_s $$

Here, ( \sigma_s ) is a symbol-specific volatility scale and ( w_{s,k} ) maps each market symbol to a macro factor. We also compute freshness, coverage, and stability for every factor snapshot.

2. Theme Detection and Theme Temperature

News is filtered through reliable-source checks, keyword-weight scoring, region tagging, asset tagging, and near-duplicate removal. Each theme gets a temperature based on velocity, mention volume, source diversity, cross-region spread, and market alignment.

$$ T = 0.35V + 0.25M + 0.20D + 0.10R + 0.10A $$

Where:

  • (V) = mention velocity
  • (M) = mention score
  • (D) = source diversity score
  • (R) = regional spread score
  • (A) = market alignment score

Theme lifecycle is then classified with hysteresis into states like hot, warming, neutral, cooling, and cold.

3. Model-Assisted Theme Ranking

For briefing quality, Atlas adds a lightweight online scoring layer:

  • StandardScaler for feature normalization
  • synthetic data augmentation when the live batch is small
  • SVD to derive a latent importance target
  • GradientBoostingRegressor for importance
  • RandomForestRegressor for temperature and market confirmation
  • KMeans for outlook-state clustering

This gives us explainable ranking without turning the product into a black box.

4. Directed Shock Propagation

Scenario Lab uses a directed weighted graph built with NetworkX. A scenario shock starts in an origin region and propagates along transmission edges with driver-specific multipliers and step decay.

$$ \phi_{u \to v}^{(t)} = I_u^{(t-1)} \cdot b_{uv} \cdot m_{uv}(d) \cdot \delta^{\,t-1} $$

$$ I_v^{(t)} = \mathrm{clamp}\left(I_v^{(t-1)} + \phi_{u \to v}^{(t)}, 0, 1\right) $$

This lets Atlas simulate how a rate shock, oil spike, FX event, or geopolitical conflict spreads across regions and asset classes.

5. Risk Aggregation

Risk Radar converts factor pressure into category scores using weighted sums, then derives higher-level desk metrics such as overall risk, systemic stress, narrative divergence, and tail-risk premium.

$$ R_c = \mathrm{clamp}\left(50 + 10.5 \cdot \sum_i \alpha_{c,i} f_i,\; 0,\; 100\right) $$

This makes the risk surface interpretable and directly tied to observed factor moves.

6. Historical Analogue Retrieval

Atlas compares the current factor vector with curated historical regimes using weighted cosine similarity:

$$ \mathrm{sim}(x, y) = \frac{(w \odot x) \cdot (w \odot y)}{|w \odot x| \, |w \odot y|} $$

This helps the user answer: what past regime looks most like today, and what followed next?

7. Confidence Scoring

Every major surface returns a confidence score computed from data freshness, coverage, and stability:

$$ C = 100 \cdot (0.35F + 0.40G + 0.25S) $$

This keeps the system honest about when evidence is strong versus when the desk should stay cautious.

Why This Matters

Most macro tools either provide raw data without interpretation or AI text without defensible reasoning. Atlas bridges that gap. It uses deterministic market math, graph-based propagation, historical retrieval, and model-assisted ranking to produce outputs that are fast, useful, and explainable enough for real financial decision-making.

Built With

Share this project:

Updates

Submission history