How predictions are made

The prediction loop

PrepareEvidence, dates, past outcomes
PredictThree views and a Combined call
SaveCheck the target and record the run
GradeCompare with the actual close
ReviewCarry lessons into the next run

A run covers one ticker at four horizons: a week, a month, a quarter, and a year. Each lens can provide a target price, expected return, direction, scenario range, rationale, and conviction for each horizon. It can also abstain when the evidence does not support a call.

The AI interprets evidence and proposes the forecast. MeebleAI supplies the price anchor, calculates the baseline and numerical checks, stores the result, and grades it later. A convincing explanation does not determine the grade.

Two ways to run the same framework

In your AI app, using your subscription: The selected model follows MeebleAI’s prepared brief through the connector. Relevant finance skills and calculation tools can help when the app actually exposes them. The shared instructions cover Claude, ChatGPT/Codex, Gemini, Grok, and Mistral; connector and tool availability varies by app. Installing an add-on alone does not prove that a particular run used it. Model identity and capability use are reported by the app, rather than independently verified execution traces. The workflow stops at the app’s allowance and has no paid inference fallback.
Through MeebleAI’s API-based runs: MeebleAI orchestrates separate lens tasks and a Combined strategist using the configured provider and models. These runs use your API key or available MeebleAI credit. The first three tasks receive their own evidence, without the earlier lens answers; the Combined task receives their results. A plugin installed in your chat app is not automatically available to this path.

Where the evidence comes from

Prices and technical indicators: Tiingo supplies daily market prices, including the anchor used by subscription runs. MeebleAI derives recent moves, trading ranges, volatility, moving averages, and momentum indicators from daily history. This is end-of-day data: during a trading session, the latest available close can be the previous session’s close. It is not a live quote.
Company financials: SEC EDGAR filings supply company financial facts. Price-based ratios can combine those facts with Tiingo prices. A filing’s reporting period and publication date matter: a recently retrieved fact is not necessarily a recent business result. Coverage can be incomplete, particularly for unsupported securities or ETFs.
News: Public RSS feeds include Yahoo Finance, Google News, Reuters, and MarketWatch. MeebleAI filters and deduplicates headlines before analysis, but feeds can still contain stale, repeated, or irrelevant items. A headline’s presence does not make it reliable evidence for that stock or horizon.
Optional supplemental research: When a subscription run’s connector supports research registration, the app can submit public sources with their URLs, summaries, lens, and known timestamps before drafting targets. MeebleAI freezes that research into the run’s evidence checkpoint. It checks the submission’s structure, not the truth of every source. If registration is unavailable, the run uses supplied evidence only. Additional research cannot replace the anchor price, target dates, or grading rules.

Runs keep their evidence mode so the Combined learning calculations can distinguish supplied-only forecasts from those with registered research. Missing dates stay unknown. Private portfolio or account information is not part of this research workflow.

The four lenses

A lens is a perspective on the same stock. These perspectives can share a model, overlapping facts, and the same biases; agreement between them is not four independent confirmations. Each supported lens is recorded and graded separately.

Technical: what is the price doing?

This lens reads price history: recent returns, trading ranges, annualized volatility, 20- and 50-day simple moving averages, and the 14-day relative strength index (RSI), where available. Moving averages describe trend; RSI describes recent momentum. Neither is a guaranteed trading signal.

It connects those observations to a numerical target and a plausible range, explaining where a trend could persist or reverse. Volatility should constrain the scenario. A short history or a missing indicator should reduce the strength of the argument; a weekly pattern alone may not support a one-year forecast. News and business valuation are reserved for their own lenses.

What a review examines: whether the call extrapolated momentum too far, missed a reversal, or underestimated the size of the move at that horizon.

Fundamental: what can the business support?

This lens considers valuation, earnings, growth, profitability, the balance sheet, and cash flow where the evidence provides them. The finance workflow asks for an explicit bridge from business assumptions to price, such as earnings per share multiplied by an assumed valuation multiple, or a cash-flow scenario. Reported facts and model estimates should be distinguishable.

The model should test how changing a key assumption changes the target, and use peer comparisons only when the evidence supports them. A good long-term valuation argument may offer little evidence about next week. Unavailable financials can make the entire lens inapplicable.

What a review examines: whether growth, margins, cash flow, or the valuation multiple justified the target—and which assumptions remained unverified.

News: what changed, and for how long?

This lens interprets supplied headlines, summaries, sources, and publication dates. The finance workflow asks whether an event is new, relevant to this company, likely to matter at the chosen horizon, and already reflected in the price. Several articles about one event should be treated as one catalyst, rather than several independent signals.

The target needs a reason for both the direction and the size of the proposed move. A positive headline is not by itself a valuation model. Unknown dates, weak relevance, or a short-lived event can justify low conviction or abstention, including abstaining across all four horizons.

What a review examines: whether the news was already priced in, whether duplicate coverage exaggerated its importance, and whether its impact lasted as long as the forecast assumed. A price miss alone cannot establish the event’s cause.

Combined: what survives the comparison?

MeebleAI first excludes unsupported lens targets for each horizon. It calculates an eligible-lens baseline using the starting weights below, then adjusts those weights when enough comparable graded history exists. Remaining weights are normalized to total 100%; an abstention is not a zero-return forecast.

Starting weights before historical adjustment
Horizon	Technical	Fundamental	News
Week	45%	20%	35%
Month	35%	40%	25%
Quarter	25%	55%	20%
Year	20%	60%	20%

These are design choices, not measured probabilities. The Combined model starts from the calculated baseline and weighs disagreement, assumptions, counterarguments, and invalidation conditions. A different final target needs a complete explanation tied to registered evidence. If that explanation is missing or invalid, the application retains the baseline and clears confidence attached to the rejected target. With no eligible evidence, there is no supported Combined forecast.

What a review examines: whether combining the lenses helped, whether a discretionary adjustment improved on the frozen baseline, and how much the conclusion depended on a single lens.

What confidence means

An 80% confidence reading means strong model conviction in that particular target and thesis. It does not mean the prediction succeeds 80% of the time.

The current rubric asks the model to weigh evidence strength and freshness, conflicting evidence, sensitivity to assumptions, and uncertainty over the horizon. The score is a model judgment, not a calculated success probability, historical hit rate, or promise that a scenario range will contain the outcome.

Below 40%: Weak conviction.
40% to below 70%: Mixed or moderate conviction.
70% to below 90%: Strong conviction.
90% to 100%: Exceptionally strong; needs explicit justification.

These bands are qualitative anchors. Current subscription submissions require an explanation for every non-empty confidence score, naming the strongest support and the main remaining uncertainty. Unknown confidence is left empty. More tools, a prestigious model, or agreement among lenses do not earn a higher score by themselves.

Confidence belongs to the specific target: when a numerical repair or guardrail changes it, the old confidence may be cleared. Older forecasts keep their original confidence meaning; they are not retrospectively relabeled under this rubric.

Checks before saving

MeebleAI checks that the result matches the run’s required structure and evidence checkpoint. It reconciles target prices, expected returns, direction, and range ordering; checks evidence references and lens availability; and enforces the Combined decision rules. Subscription results are bound to the prepared evidence by a digest, so a submission for a different checkpoint is rejected.

These checks can reject or adjust a result. The final checked targets and confidence are the ones that should be saved. Passing validation means the result satisfies those checks; it does not certify the reasoning or predict investment success. A validated draft or a registered research checkpoint is not yet a published run.

Supported predictions are saved with their horizons and grading policy before the outcome is known. An abstention remains visible as missing support, rather than being silently replaced by a flat call. Older interpolated horizons are treated separately in the newer integrity and learning calculations.

How predictions are graded

Target dates start at 7, 30, 90, and 365 calendar days after the run date. A weekend target moves back to Friday. When a horizon is due, the evaluator uses the first available daily close on or after its saved target date, which also handles a market holiday. If market data is not yet available, grading waits. The starting price can be rebased for a material adjustment in the historical price series, such as a split.

Direction and price error answer different questions

Direction: did it call up, down, or flat correctly?: Small moves count as flat. The band is ±1% for a week and grows with the square root of the horizon: about ±2.07% for a month, ±3.59% for a quarter, and ±7.22% for a year. The predicted and actual direction must match.
Price error: how far was the predicted return from the outcome?: MeebleAI measures the absolute difference between predicted and realized returns in percentage points. The price grade is correct when that error is within the saved tolerance. The base tolerances are 1.5 points for a week, 3 for a month, 6 for a quarter, and 12 for a year, scaled for the stock’s volatility.

Exact calculations and the role of ranges

Expected return = (target price ÷ anchor price − 1) × 100.
Realized return = (actual close ÷ grading anchor − 1) × 100.
Price error = absolute value of (expected return − realized return).
Flat band = 1% × square root of (horizon days ÷ 7), with the boundary included as flat.
Tolerance multiplier = annualized volatility ÷ 25%, limited to a minimum of 0.5× and maximum of 2.5×. Missing usable volatility means the base tolerance.

For current forecasts, the volatility input and grading thresholds are frozen before prediction. The model’s low/high scenario range does not set its grading tolerance, so widening a range cannot make that price grade easier to pass. A scenario range is not a calibrated statistical confidence interval. Legacy forecasts without a saved policy use their older evaluation path.

Every supported lens and horizon receives its own outcome. A missing target or an abstention has no price call to grade and is not counted as a correct forecast.

How the next run learns

Learning here means changing the evidence, review context, and blend used in a future run. MeebleAI does not retrain the underlying model’s weights. There are several distinct feedback paths, and they do not all run in every mode.

Subscription runs: outcomes and review notes

During preparation, MeebleAI scores due predictions without an LLM or embedding call.
The brief includes recent graded outcomes for that account and ticker, up to 12 existing stored lessons, and review notes from the last three published subscription runs.
The selected model can write new review notes tied to the exact supplied prediction and horizon. Those notes are optional, saved with the run, and can inform a later brief.
Comparable numerical outcomes also feed the Combined baseline described below.

Subscription grading does not secretly purchase an API critique or create new vector-search lessons. A successful run with no review notes still adds forecasts that can later be scored, but it has not created a written lesson by itself.

API runs: stored critiques, retrieval, and bias context

The evaluator scores the outcome. Where the configured funding and model call succeed, it also reviews the forecast’s original reasoning and inputs to write a short critique for that lens and horizon. A critique can be written for a correct call as well as a miss. Failed critique generation does not prevent the numerical grade.
Critiques are stored as lessons with searchable vector representations. Future live runs retrieve similar lessons from the owner’s lesson library, normally up to two per horizon, with a broader fallback when there are no horizon matches.
Technical, Fundamental, and News receive lessons tagged for their lens, plus applicable untagged older lessons. Combined can receive lessons across lenses. The selected lesson IDs are recorded for later measurement.
With at least five comparable graded observations, a lens can receive a warning about historical over- or underestimation. The prompt includes it when average signed error is at least one percentage point. This is context for judgment, not an automatic subtraction from the next target.

The live lesson library can include historical simulation lessons, and semantic retrieval can return lessons from other tickers. That broader context differs from the strictly matched live outcomes used for the Combined weights. Historical replay uses its own time cutoffs to avoid retrieving outcomes that were not yet available.

Combined weights: learn only from comparable outcomes

For each horizon, the calculation matches ticker, account, provider, model, evidence mode, and the applicable workflow version. It uses outcomes available before the new forecast, separates live runs from simulations, and excludes unsupported or synthesized lens targets. Comparisons use the same eligible lenses from complete matched runs, with at most one run per forecast date and at most 252 dates.

Fewer than 20 matched dates, or fewer than two eligible lenses, leaves the starting weights unchanged. With enough history, price-grade hit rates are smoothed toward the pooled average and blended with the starting weights. Historical evidence can contribute at most 75% to that weight update. Equal rates also leave the defaults unchanged. Confidence scores do not determine these learned weights.

The weight update, step by step

For n matched forecast dates and each eligible lens:

Calculate the pooled price-grade hit rate across those lenses and dates.
Smoothed rate = (lens hits + 20 × pooled hit rate) ÷ (n + 20).
Normalize the smoothed rates to weights totaling 100%.
Historical share = the smaller of 0.75 and n ÷ (n + 50).
Final weight = (1 − historical share) × starting weight + historical share × learned weight.

Candidate history is bounded to three years and 6,000 rows per horizon before matching and the 252-date limit. This reduces the chance that repeated runs on one day or incomparable populations dominate the blend.

Check whether the feedback helped

When a current Combined forecast matures, MeebleAI also grades its frozen equal-weight, starting-weight, and adaptive baselines against the same outcome. That allows comparison of the final strategist call with the alternatives it had at prediction time.

For API forecasts with recorded lesson use and enough comparable earlier history, the system compares the new absolute error with the historical average error available at prediction time. Lessons associated with fewer improvements can be deprioritized in live retrieval after enough observations. These are observational comparisons: market conditions, model differences, and selection effects can explain a change. They do not establish that a lesson caused better predictions.

How to read the results

Read the target with its horizon and assumptions. A long-term valuation case and a weekly momentum call answer different questions.
Look at the sample and coverage. A high hit rate on a few graded calls says little. Abstentions and pending outcomes matter alongside completed grades.
Separate conviction from performance. Compare realized direction and price error. Confidence is the model’s judgment about the current thesis.
Check the evidence mode and model. More research or a finance skill can change the process; it does not prove a more accurate forecast. Different models can use the same conviction scale differently.
Keep simulations separate. Public track-record statistics use live predictions. Historical simulations are a different exercise; limited archived news and the model’s prior knowledge of history can affect their interpretation.
Expect uncertainty. Models can misread facts or produce persuasive but weak reasoning. New information and market shocks can overturn a supported thesis. The learning loop is a mechanism to evaluate, not a guarantee of improvement.

MeebleAI is an investing playground for examining AI opinions and their results. Predictions are impersonal research, not recommendations tailored to your situation.

See the public track record Explore predictions Read the data disclosures