Consensus score

Consensus starts after individual AI Agent disagreement is preserved

The iPulse Consensus Score compresses multiple advisor forecasts into a bounded signal, while still accounting for return magnitude, disagreement, path volatility, dividends, and economic guardrails.

Updated July 3, 2026

Browse AI Agents Forecasting Top Picks

Overview

A score is a summary, not a guarantee

iPulse first produces structured forecasts from individual AI Agents. The Consensus Score then summarizes the direction and strength of the grouped forecasts into a range from -1000 to +1000. Positive scores indicate bullish risk-adjusted consensus; negative scores indicate bearish risk-adjusted consensus.

The score is designed for educational market research and comparison. It is not personalized financial, investment, tax, or legal advice.

Inputs

The score listens to more than average return

A simple average can overstate conviction when forecasts disagree or when the forecast path is unstable. iPulse therefore considers return direction, return magnitude, cash hurdle, dividend-inclusive total return, historical volatility, forecast dispersion, path volatility, and guardrail rules.

The formula deep dive below appears after prompt assembly in the docs flow because the score only makes sense once the AI Agent, mode, prompt, context, and output schema pipeline is understood.

What improved

The formula changed after ranking edge cases exposed weak assumptions

Consensus scoring has been refined through operating experience. The main lesson is that a clean-looking formula can still rank assets badly if it hides return magnitude, over-penalizes normal disagreement, or lets small low-volatility moves dominate larger economic outcomes.

Pure Sharpe-style ranking was too aggressive

Dividing everything by historical volatility can make very different return forecasts look similarly important. iPulse now blends volatility-aware signal with a return-anchor component so magnitude remains visible.

Dispersion should reduce confidence, not erase signal

Advisor disagreement is informative. The score discounts excess return when forecasts disagree, but caps that discount so a real directional signal is not zeroed out mechanically.

Volatility penalties need a dead zone

Small path variation should not punish a forecast when agents broadly agree. The current approach uses a non-linear penalty with a return-sensitive dead zone so only meaningful path instability drags the score down.

Positive return is not automatically a buy

The score uses cash-hurdle and spread-quality guards so low positive returns or fragile upside can remain neutral. This keeps public labels more conservative and easier to defend.

iPulse Consensus Score -- Formula and Methodology

How the ranking signal is computed

The score is not a promise. It is a disciplined compression of return direction, return magnitude, dividend income, historical volatility, analyst agreement, forecast-path stability, and economic guardrails into a single index from -1000 to +1000.

Current scoring version

v7.3 uses a 3.6% cash hurdle, capped MAD discount, 35%/65% signal blend, high-volatility return-anchor scaling, buy-side safety checks, and a sell-side economic floor.

Simple-language version

1. Start with each AI Advisor's forecast. iPulse asks what return the prediction implies over the selected horizon.
2. Add dividends into the canonical return path. For dividend-paying assets, expected income is folded into total return before ranking; price-only comparison is optional when ex-div diagnostics exist.
3. Put every forecast on the same annual scale. A 1-year forecast and a 5-year forecast can be compared without mixing timeframes.
4. Measure excess return above cash. The first hurdle is 3.6%, so low positive returns do not automatically become bullish.
5. Discount fragile consensus. Return MAD reduces excess-return magnitude before scoring, capped at 60% of excess return.
6. Balance risk and return size. The score blends historical-volatility awareness with a return-anchor term that scales down very high-volatility assets.
7. Reward agreement and penalize messy paths. Consistency can lift clean consensus; the volatility penalty dampens unstable forecast paths.
8. Keep signals economically honest. Weak bullish returns can be downgraded to NEUTRAL, and negative terminal economics stay visible as at least PARTIALLY SELL.

Convert each prediction into an annualized total-return voice

Input: One AI prediction's price path, forecast horizon, and dividend yield estimate (when applicable).

R_adj,i = (((1 + R_base,i/100) * (1 + y)^Y) - 1) * 100
R_i = (((1 + R_adj,i/100)^1/Y) - 1) * 100

Output to next step

R, the annualized total-return voice for one prediction.

Group voices so one source cannot drown out the rest

Input: All annualized voice returns for the asset, grouped by selected mode: prediction, AI Advisor, or AI Persona.

R_g = mean(R_i in group g)

Output to next step

R_g, the grouped annualized return for each independent voice bucket.

Discount return spread and measure excess return over cash

Input: Consensus annualized return, cash hurdle, and return MAD across grouped voices.

E = R - 3.6%
R_conf = sign(E) * max(0, |E| - min(0.35 * MAD, 0.60 * |E|))

Output to next step

R_conf, the confidence-adjusted excess return used for direction scoring.

Blend risk-aware return with return magnitude

Input: R_conf plus the asset's historical volatility and high-volatility return-anchor multiplier.

S = 0.35 * R_conf / max(sigma_hist, 5%) + 0.65 * R_conf / (15% * M_vol)

Output to next step

S, the normalized direction signal before tanh bounding.

Turn direction into consensus strength and agreement

Input: The consensus signal S and the distribution of grouped voice signals.

b = tanh(S / 2)
consistency = e^-sigmaS

Output to next step

b, the bounded direction; consistency, the agreement multiplier.

Dampen unstable forecast paths only after a proportional dead zone

Input: Consensus return magnitude and cross-voice forecast dispersion.

deadZone = clamp(0.5 * |R|, 1%, 10%)
volatilityPenalty = 0 when sigma <= deadZone
volatilityPenalty = min(1, ((sigma - dz) / (15 - dz))²) * 0.15

Output to next step

volatilityPenalty, the instability discount applied to the final score.

Calculate the final score, then apply economic guardrails

Input: Direction, consistency, volatility penalty, annualized return, and return spread quality.

Score = round(b * (0.85 + 0.15 * consistency) * (1 - volatilityPenalty) * 1000)
BUY guard: R < 3.6% or (R - MAD) < -1% => NEUTRAL
Sell guard: R < -0.5% => at least PARTIALLY SELL

Output to next step

A -1000 to +1000 iPulse score and final rating band.

Symbol dictionary

R_base,i: price-only compounded return for one prediction voice from the generation anchor price
y: annual net dividend yield, estimated from 3-year historical cash-dividend events when usable
Y: forecast horizon in years
R: annualized total return after dividend uplift (or zero uplift when no usable dividend history exists)
E: excess return above the 3.6% cash hurdle
R_conf: confidence-adjusted excess return after the capped MAD discount
MAD: mean absolute dispersion of grouped annualized returns, used to discount fragile upside
sigma_hist: 10-year historical annualized volatility, floored at 5%
M_vol: high-volatility denominator multiplier, 1.0x to 2.0x
alpha: 0.35 Sharpe weight; 0.65 return-anchor weight
b: bounded direction signal after tanh compression with k=2
sell guard: minimum sell-side visibility when annualized economic return is below -0.5%

Dividend handling

Leaderboard ranking uses dividend-inclusive total return as the canonical basis. Price-only comparison is shown only when ex-div diagnostic fields are available for that snapshot.

The yield is estimated from recent cash-dividend history, usually a 3-year window, using dividend amounts relative to prior close prices. Assets with no usable dividend history receive no dividend uplift.

Dividend uplift is most meaningful for 1-year and longer horizons. It can change both displayed annualized return and ranking score because income is treated as part of total shareholder return.

Agreement, dispersion, and recompute shocks

Consistency measures how tightly grouped the AI Advisor signals are after voice grouping. A narrow signal distribution receives a higher consistency term; a wide distribution reduces conviction.

Small deviations are ignored. Only genuinely unstable forecast paths reduce the score in a meaningful way.

Recomputed leaderboard snapshots compare the current market price with the prediction path. Post-analysis shocks below 12% are not dampened. Larger shocks attenuate the score input toward the 3.6% cash hurdle with capped horizon weights of 0.50 for 1Y, 0.30 for 3Y, and 0.20 for 5Y. Close prices and displayed return fields remain raw market/mechanical values. The raw pre-shock return is still used for the sell-side guard, so a below-current terminal forecast cannot hide as NEUTRAL.

One Score per AI Prediction

Every individual prediction counts as an equal vote. This is the raw, unadjusted signal.

One Score per AI Advisor

Predictions from the same AI Advisor are averaged before scoring, so one advisor cannot dominate simply by producing more forecasts.

One Score per AI Persona

Predictions sharing the same analytical persona are averaged into one vote per perspective. This is the most diversity-aware view.

Performance-weighted mode

Reserved for later use once enough outcome history exists to weight advisor votes by measured forecasting performance.

Rating bands

STRONG BUY

> +300

BUY

+131 to +300

NEUTRAL

-30 to +130

PARTIALLY SELL

-200 to -31

SELL ALL

< -200

BUY and STRONG BUY are downgraded to NEUTRAL when annualized expected return is below the 3.6% cash hurdle or when the lower return spread (return minus MAD) is below -1%. Sell-side visibility is protected: if the full-horizon annualized economic return is below -0.5%, the final rating is forced to at least PARTIALLY SELL even when volatility or shock dampening would otherwise soften the score toward NEUTRAL.

Formula Version is stored with the leaderboard output so every rank can be traced back to the exact scoring specification used at generation time. Different versions can change thresholds, weighting, volatility handling, dividend treatment, or diagnostic logic. Production leaderboard ranking uses the canonical dividend-inclusive score and signal fields.

Limits

How to read the score responsibly

Consensus is useful, but disagreement is also information. A high score should still be read with the supporting AI Agent reports, forecast horizon, asset volatility, thesis risks, and scenario assumptions.

Eligible signed-in users can open Config Details on advisor reports to inspect which model, mode, task, and prompt assembly were used for a given prediction.

Read Config Details tutorial