The AlphaLucid Analysis Guide

Everything the platform computes — explained in plain words, with the underlying math wherever numbers appear.

1Core concepts

The building blocks that show up everywhere.

Price & Change %

The price is the last quote. The daily change compares the current price with yesterday's close.

Return & annualized return

The return over a period = how much the value grew. Annualized = the equivalent "per year" rate, so you can compare different periods.

Percentile

Tells you where you rank within a group, from 0 to 100. The 80th percentile = "better than 80% of the comparison peers". It's the key mechanism behind the AlphaLucid Score.

For metrics where "lower is better" (e.g. P/E), the platform inverts the percentile (100 − p), so a high score always means "good".

2The AlphaLucid Score centerpiece

A single 0–100 number with the "why" visible: an explainable grade, not a black box.

The idea

Instead of absolute thresholds ("P/E under 15 = cheap"), the platform compares each instrument with its peers (the instrument universe) and places it on a 0–100 scale via percentiles. That way the score adapts to market conditions.

The 6 dimensions

How the number is reached (step by step)

1) For each of the ~30 factors, the instrument's percentile vs the universe is computed. 2) The factor percentiles within a dimension are weight-averaged → a sub-score (0–100). 3) The 6 sub-scores are weight-averaged → the overall score.

The rating

The score maps to a plain-language RATING of its strength — a description of the analysis, not buy/sell advice:

The confidence band

The score comes with a band (e.g. "64 ± 6"). The more complete and consistent the data, the tighter the band. This is honest uncertainty: it won't promise false certainty.

3Fundamentals & ratios

A company's "vital signs". All are real (from the data provider), TTM = the last 12 months.

Valuation (how expensive)

Profitability (quality)

Financial health

Growth & dividend

4DCF valuation — fair value

DCF = Discounted Cash Flow. It answers "what should it be worth, given the cash it will generate?"

The idea

Money in the future is worth less today ($1000 in 10 years < $1000 now). DCF estimates future cash flows and "brings them to the present" with a discount rate r (cost of capital / risk).

FCF = Free Cash Flow; r = discount rate; g = long-term growth; t = the year; n = the last projected year.

Variants by business type

The platform picks the right model automatically: FCF (normal companies), Excess Returns (banks/financials), AFFO (REITs / real estate), NAV (ETFs, the value of the underlying assets).

Margin of Safety

The percentage gap between the estimated fair value and the current price. Positive = potentially undervalued.

Tornado (sensitivity analysis)

Shows which assumption moves the value the most (growth, margin, discount rate). Each bar in the "tornado" is how much the value changes if one assumption varies by ± a little. It helps you see where the biggest uncertainty is.

5Probabilistic scenarios (Monte Carlo)

Instead of a single target, a range of possible outcomes with their probabilities.

How it works

Thousands of price "futures" are simulated from an expected return (drift) and a volatility (how "jumpy" it is). The per-step model is geometric Brownian motion:

μ = expected return; σ = volatility; t = horizon (years); Z = a random shock (normal distribution). Repeat thousands of times → a distribution of outcomes.

How the inputs are estimated (honestly)

The quality of the output depends entirely on μ and σ, so the platform estimates them carefully rather than naively:

• Volatility σ = the stock's realized volatility from ~2 years of real daily returns (not just beta), pulled toward a long-run level — calm or panic don't last 5 years (mean reversion).
• Drift μ = a forward-looking expected return (dividend yield + capped earnings growth, Bogle-style), pulled toward a long-run market anchor (~8%) — NOT last year's trailing return extrapolated forward. Scenarios are centered on this drift.
• A one-year floor keeps a single year from "losing more than 100%".

This rework fixed pathological outputs the naive model produced (e.g. a bogus ~99% chance of loss on a healthy mega-cap). The model lives in one place and is shared by the app and the calibration tools, so what you see is what we test.

Selectable horizon

You can run the analysis over 1M, 3M, 6M, 1Y, 5Y, 10Y, 15Y or 20Y. The horizon drives both the historical lookback and the forward projection, so short- and long-term views are consistent.

What you read

6Technical analysis

Signals from the price chart (not the fundamentals). On real data.

Moving averages (SMA) & trend

SMA = Simple Moving Average (the simple mean of the last N prices). Compare the 50-day vs 200-day SMA for direction.

RSI (Relative Strength Index)

Measures how "overbought/oversold" something is, on a 0–100 scale, over 14 days. Above 70 = overbought; below 30 = oversold.

Support / Resistance & breakout

Support = a recent low (a "floor"); Resistance = a recent high (a "ceiling"). If the price breaks above resistance → Breakout; if it falls below support → Breakdown.

7Trained ML model

A real statistical model (logistic regression), trained on price history — with no "look-ahead" (it can't see the future while learning).

What it does

It estimates the probability that the instrument rises over the next ~month, from price signals known at that point in time.

The features

How it "decides" (the math)

Each feature is standardized (z-score), combined linearly with learned weights, then passed through a sigmoid that maps to a 0–1 probability.

The weights (w) are learned by gradient descent over thousands of historical examples (at each past date: the features then → did it rise over the next 21 days or not).

Validated honestly

The model is trained only on instruments with real data, with non-overlapping samples (so a label can't leak into the training window), and it is evaluated out-of-sample — on dates it never trained on. The "training accuracy" shown is in-sample and optimistic by nature; the honest measure is the out-of-sample evaluation and the public Track Record. We are transparent that the edge of a price-only model is modest and uncertain — the platform won't oversell it.

8Sentiment / NLP

NLP = Natural Language Processing. It reads real news headlines and assigns a tone (positive/negative).

How it's computed

Each headline is scanned with a financial lexicon (lists of positive words: "beat, surge, upgrade…" and negative: "miss, plunge, downgrade…"). A headline's score is the balance of words:

It also handles negation ("not strong" becomes negative). The "delta" shows how the tone changed vs the previous reading.

9Backtesting

Tests a strategy on real historical prices, honestly — with costs and without "cheating" with the future.

The momentum strategy (the platform's example)

At each rebalance (monthly): rank the instruments by past momentum, hold the top K, measure the next month's return, apply a trading cost, and compound. Compare with an "equal-weight" benchmark (all of them, in equal weights).

"No look-ahead" = at each moment it uses ONLY the information available then. That's what makes the backtest credible (it also shows when the strategy doesn't beat the market).

10Portfolio — return & risk

TWR — Time-Weighted Return

The "pure" return of the investments, which removes the effect of your deposits/withdrawals. Good for judging selection, not how much you put in.

XIRR — money-weighted return (annualized)

Accounts for how much and when you invested. It's the rate r that makes all the cash flows (deposits −, final value +) cancel out on a discounted basis:

The platform caps extreme values (e.g. positions opened the same day) so it doesn't display absurd percentages.

Allocation & marginal contribution to risk

Allocation = what percentage of the portfolio each position is. Marginal risk = how much a position adds to the total portfolio risk (it also accounts for correlation with the rest) — sometimes a small position adds a lot of risk if it's highly correlated.

11Dividend safety (A–F)

A grade (A = very safe … F = risky) for how likely the dividend is to be maintained.

It looks mostly at the Payout Ratio (how much of profit is paid out — under ~60% is comfortable), earnings growth, debt and cash flow. A payout above 100% (paying out more than it earns) = a warning sign.

12Calibration & Track Record

How "honest" are the platform's probabilities? This is where it's measured.

Brier Score

Measures the accuracy of probabilistic predictions. Lower = better (0 = perfect, 0.25 = a coin flip).

Reliability curve & recalibration

It checks "when I said 70%, did it actually happen ~70% of the time?". If not, the model recalibrates (adjusts the probabilities) to be more honest. The Track Record shows raw vs recalibrated Brier. Predictions are logged and resolved publicly — we score ourselves in the open.

13AI features

A layer of intelligence on top of the numbers. (The actual AI text appears with an Anthropic key; without it, a variant derived from the score.)

RAG with citations

RAG = Retrieval-Augmented Generation. The AI answers only from the real data in the platform and puts a citation after each number. Zero invented figures.

Bull / Bear / Judge (adversarial)

Three perspectives: one builds the optimistic case (Bull), one the pessimistic case (Bear), and a "Judge" weighs both and gives a balanced verdict. It reduces bias.

Behavioral guard

When you're about to act on an emotional wave (FOMO into a big run, panic into a drop), it shows a calm message that brings you back to the right question: "has the thesis changed, or just the price?".

Self-consistency (the confidence)

The same analysis is run several times; the more the results agree, the higher the confidence (and the tighter the band).

The night analyst

A scheduled job that scans your watchlist/portfolio/theses and reports only what changed — a digest instead of noise.

Thesis-erosion monitor

It compares your original assumptions (why you bought) with current reality; if the premises have broken, it warns you.

AI (Claude) calls are the only part that costs tokens, so they're enabled per-account from the admin panel (off by default). Everything else — score, DCF, scenarios, technical, ML, sentiment — is free and open.

14World indices

An index measures a basket of stocks — the thermometer of a market.

E.g.: S&P 500 (500 large US companies), NASDAQ (tech), DAX (Germany), Nikkei 225 (Japan). Each trades in its own currency (USD, EUR, JPY…), which is why the numbers aren't directly comparable. On each index page you also get an associated ETF you can analyze in depth (e.g. S&P 500 → the VOO ETF).

15Glossary & abbreviations

AlphaLucid · Analysis Guide. The formulas reflect how the platform computes the figures it displays.