Methodology · v1

How the swarm works.

We built FutureDaily on one simple bet: a single forecasting model is wrong in interesting ways, but a small ensemble of opinionated specialists, critiqued adversarially and weighted by track record, is wrong in calibrated ways. Here's exactly what we do, every morning.

Daily, 8:00 AM IST

The five-phase pipeline

  1. step 1
    Phase 1 — Independent forecasts

    Each of seven agents reads the same macro context (rates, oil, reserves, REER, FPI flows, geopolitics) and produces a point estimate, an 80% confidence band, top three drivers, and a single load-bearing assumption. They do not see each other's work.

  2. step 2
    Phase 2 — Adversarial CIO critique

    A separate critique pass evaluates every agent's weakest assumption, names the falsifying data point, and tags the agent as overconfident, underconfident, or well-calibrated. The CIO publishes its own base case alongside.

  3. step 3
    Phase 3 — Disagreement map

    We compute the range, median, and standard deviation across agents at each horizon, identify outliers, and flag whether the disagreement is directional or magnitude.

  4. step 4
    Phase 4 — Horizon-weighted aggregation

    Each agent receives a weight per horizon, calibrated by their rolling 30-day Brier score and downweighted by the CIO critique. The consensus point is a weighted mean; bands are weighted standard deviation × 1.28 (80% coverage under normality).

  5. step 5
    Phase 5 — Trade-actionable output

    We surface the conviction level (low / medium / high), the five monthly indicators that would invalidate the call, and the swarm's collective tail risks.

Lenses, not models

The seven agents

Horizon-aware, calibration-weighted

Weighting scheme

Different lenses dominate different horizons. Technicals and policy reaction win the short term because that's when central-bank behaviour and mean reversion matter most; fundamentals and structural factors take over by year three.

Macro Fundamentals
12M 12%24M 18%36M 22%
Dominates long horizons via REER/PPP convergence.
BoP & Flows
12M 13%24M 18%36M 20%
BoP identity is anchor at all horizons.
Policy Reaction
12M 22%24M 18%36M 10%
Short-term FX = central-bank behaviour; fades at 3Y.
Technical / Quant
12M 22%24M 12%36M 5%
Mean reversion dominates 12M; fades at 3Y.
Geopolitical
12M 15%24M 18%36M 20%
Oil/regime risk critical near-term; structural at 3Y.
Historical Analogist
12M 8%24M 8%36M 13%
CIO-skeptical; more useful at longer horizons.
Contrarian / Red Team
12M 8%24M 8%36M 10%
Fat-tail protection; compounds at 36M.
The trust block

Scoring and recalibration

Every morning we publish a forecast. Every evening we compute the actual close, mark every horizon to market, and write the result to a public scorecard. The headline metric is a rolling 30-day Brier score; we also publish hit rate (in-band), directional accuracy, and average error magnitude.

Agent weights are recalibrated weekly on an exponential-decay Brier schedule — the most recent forecasts count most, but we never throw away the long tail. An agent that lands consistently outside its own 80% band does not earn weight no matter how compelling its prose.

We also publish the things that would change our mind: five monitored indicators with concrete triggers (Brent above $115 sustained, FPI outflows above $5 bn for three months, Core PCE above 2.7% through Q3). If any one breaks, we say so within the day.

The honest list

What we don't do

FAQ

Frequently asked

What is FutureDaily?

FutureDaily is an open, public daily forecasting site for USD/INR. Seven independent AI agents produce structured forecasts each morning at 08:00 IST. A calibration-weighted consensus is published with an 80% probability band, and graded the next trading day against the actual close.

How accurate are the forecasts?

Calibration is the primary success metric, not point accuracy. The 80% band should contain the next-day close 80% of the time. Every call is publicly graded and the rolling Brier score, in-band rate, and beat-consensus rate are visible on the scorecard.

How are the seven agents weighted?

Each agent has a static base weight per horizon (1-day, 12-month, 36-month). After each forecast, the CIO red-team tags each agent OVERCONFIDENT (-25%), UNDERCONFIDENT (+20%), or WELL CALIBRATED. Weights are then renormalised so they sum to 1.

Which agents are in the swarm?

Macro Fundamentalist (REER/real rates), BoP & Flow Analyst (current and capital account), Policy Reaction Modeler (RBI behavior), Technical/Quant (price action), Geopolitical/Structural (regime risk), Historical Analogist (analog matching), and Contrarian/Red Team (fat-tail protection).

What data do the agents use?

USD/INR spot from Yahoo Finance, DXY, Brent crude, India CPI and repo rate, US CPI and Fed funds target, plus curated geopolitical context. Each agent can use web_search to verify recent prints. India-specific macro inputs are updated when official prints land.

Is this investment advice?

No. FutureDaily is research output published openly. It is not investment advice, a managed product, or a signals service. Do not size positions from this data.

Can I cite FutureDaily forecasts?

Yes. Citation is welcome and encouraged. Cite as: FutureDaily Swarm (YYYY-MM-DD). USD/INR daily forecast. https://futuredaily.io/usdinr/article/YYYY-MM-DD. Structured data is available at https://futuredaily.io/api/today.

How often are forecasts published?

Daily at 08:00 IST (02:30 UTC). Grading runs daily at 17:30 IST (12:00 UTC) once the day's USD/INR close is available from Yahoo Finance.