Docs

Documentation

Start with the product you are using right now, then move into pricing, account management, and the full reference material.

Product Docs

Evaluation

Upload a CSV of predictions and actual prices, compare up to your tier limit, and rank models with FIS plus trading metrics.

Product Docs

Decision Intelligence

Analyze a raw dataset before model training, inspect health scores, and review recommended model families and preprocessing steps.

Account Docs

Subscriptions & Billing

Review your current tier, quota reset date, upgrade path, and self-service billing options in the account dashboard.

Full Reference

Use the maintained product reference below. It documents current behavior without sending browsers to a raw repository-only Markdown path.

Getting started

1. Before training

Upload a UTF-8 CSV in Decision Intelligence, confirm the target, then review observed facts, quality warnings, leakage signals, preprocessing guidance, and suggested model families.

2. Train and validate

Train models in your own environment. Use a held-out split for independent rows or time-ordered walk-forward validation for forecasting. Keep a simple baseline.

3. After training

Upload actual outcomes plus one or more forecast columns to Evaluation. Compare prediction error and decision utility; enable finance interpretations only for a suitable task.

Files, limits, and common errors

Supported format: CSV. Decision Intelligence needs at least two columns and enough observations for meaningful diagnostics. Evaluation needs an actual-outcome column, at least one numeric forecast column, and at least 10 valid paired rows.

Anonymous allowance: 5 successful runs per rolling week, shared across public tools. Signed-in Free accounts receive 10 evaluation units per month; each forecast candidate consumes one unit.

Common errors: verify the delimiter, remove duplicate header names, choose the correct target/actual column, and ensure numeric cells are not mostly empty. Files that are not CSV are rejected before analysis.

See Pricing for the canonical public plan matrix.

Validation & Methodology

Dataset facts and deterministic statistics are reported separately from heuristic or model-assisted recommendations. Recommendations are starting hypotheses, not validated performance claims.

For time series, preserve temporal order and prefer rolling or walk-forward validation. For independent tabular observations, use a held-out test set and cross-validation that respects groups where relevant. Always compare against a naive or simple baseline.

FIS and CER are proprietary metrics that add decision-oriented perspectives; MAE, RMSE, MASE, and related diagnostics measure different properties. No single metric establishes deployment readiness. Trading simulations depend on position rules, costs, slippage, sampling frequency, and the absence of leakage.

Limitations include small samples, unstable regimes, target leakage, selection bias, distribution shift, and uncertain real-world costs. Quantsynth does not train or deploy models and does not provide financial advice.

FIS and CER: overview and motivation

Traditional forecast-error metrics describe numerical differences between predictions and observations. They are essential, but numerical closeness does not always answer whether a forecast is useful for a particular decision. QuantSynth therefore presents FIS and CER as additional, proprietary perspectives rather than replacements for established error metrics.

The intended question is comparative: when candidate forecasts make different tradeoffs, do decision-oriented and error-oriented views rank them differently? A disagreement is a prompt for investigation, not proof that one metric or model is universally correct.

Interactive guide

Annotated FIS and CER comparison

Switch between reproducible, precomputed fixtures generated by the current backend implementation to see why conventional error metrics and decision-oriented views can disagree. These are educational fixtures, not results from your uploaded data.

Annotated output

Timing tradeoff

Precomputed backend fixture
Illustrative forecast versus actual chart An actual series and two forecast series with an annotation describing the selected tradeoff. time level
ActualForecast AlphaForecast Beta
Comparison

What changes?

Illustrative metric comparison
ForecastMAEFISCERRank cue

Interpretation guardrail: the chart and values demonstrate how to read a comparison, not how the proprietary calculations are constructed. In a real evaluation, verify the data, validation design, conventional errors, and application-specific outcomes together.

FIS explained

Decision-oriented forecast evaluation

FIS is QuantSynth's proprietary metric designed to evaluate forecast behavior from a decision-oriented perspective. It is intended to help identify candidates whose behavior may be more relevant to an application's decisions even when another candidate has lower average numerical error.

Higher FIS scores generally indicate a more favorable result under the metric's evaluation framework. Interpret the score comparatively within a consistent dataset, horizon, validation design, and application. A high score is not a probability of success, an estimate of financial return, or a guarantee of future performance.

Use FIS beside MAE, MSE, RMSE, MASE, calibration checks, baselines, and domain-specific outcomes. Do not use it as an automatic deployment rule.

CER explained

Complementary forecast evaluation

CER is QuantSynth's proprietary metric for comparing forecasts from predictive and decision-oriented perspectives within a complementary framework. Its role is to add context when candidates differ in numerical accuracy and practical relevance.

Use CER rankings as another diagnostic view over the same candidate set. FIS and CER can favor different forecasts because they are designed to emphasize different evaluation perspectives. A CER ranking is not universally superior to an FIS, MSE, or MASE ranking.

CER should be interpreted with the product's comparison table, traditional error diagnostics, the validation protocol, and the intended application. Avoid comparing isolated values produced under materially different datasets or evaluation settings.

Comparison with forecast-error metrics

MAE, MSE, and RMSE

Describe error magnitude, with different sensitivity to large errors. They directly answer numerical-closeness questions but do not by themselves encode an application's decision consequences.

MASE and scaled diagnostics

Support scale-aware comparison and baseline-relative interpretation. They remain important checks even when a decision-oriented metric ranks a candidate highly.

FIS and CER

Add proprietary decision-oriented comparison views. They do not establish causality, profitability, calibration, robustness to distribution shift, or deployment readiness.

Interpreting results

Agreement: when FIS, CER, and conventional errors favor the same candidate, the evidence is directionally consistent, but independent validation is still required.

Disagreement: inspect the forecast paths, error distribution, timing, stability across folds or windows, and application-specific outcomes. Metric disagreement is information about a tradeoff, not permission to ignore an unfavorable diagnostic.

Absolute values: avoid universal cutoffs. Compare candidates under the same preprocessing, sample, horizon, and validation protocol, and confirm that rankings persist across reasonable design choices.

Worked conceptual examples

Low error, weak timing

A smooth forecast may remain numerically close to the observed series while reacting too late for a timing-sensitive decision. Error metrics may favor it, while a decision-oriented view may prefer another candidate. This is a hypothetical pattern, not a performance claim.

Useful direction, rough magnitude

A candidate may identify relevant direction or event timing while misstating magnitude. FIS may view that behavior more favorably, while MAE or MASE exposes the magnitude weakness. Both findings matter.

FIS and CER disagree

One candidate may look stronger under FIS while another has a more favorable CER profile. Review conventional errors and application outcomes rather than treating either ranking as automatically decisive.

Experimental validation and evidence

Design: internal controlled model-selection experiments used public yfinance price histories requested from January 2018 through June 2026 for US equities, US Treasuries, gold, oil, Bitcoin, and EUR/USD. Eight synthetic or oracle-informed forecast candidates were evaluated over 63-observation rolling windows. FIS and CER selectors were compared with MSE and MASE selectors.

Dependence checks: the primary run used 192 asset-window comparisons with overlapping windows. Inference included moving-block bootstrap, Newey-West, multiple-testing adjustment, sign/permutation, rank, and asset-blocked checks. A separate 192-comparison run used non-overlapping windows.

FIS finding: in this controlled candidate set, FIS selected the known utility-leading candidate in 88.5% of overlapping windows and 84.9% of non-overlapping windows. MSE and MASE each did so in about 0.5% of windows. The FIS advantage survived the experiment's adjusted window, dependence, rank, and asset-blocked checks.

CER finding: CER produced favorable return and rank comparisons against MSE and MASE in this framework, but its known-oracle selection rate was about 1.0% and did not establish an adjusted oracle-selection advantage. The evidence for CER is therefore materially weaker than the evidence for FIS in these tested selection scenarios.

Important negative result: an adversarial sparse-large-event construction received a relatively favorable FIS result while capturing only a small share of the reference outcome in most windows. This demonstrates that FIS can overvalue sparse directional success and must be checked against magnitude, event coverage, and application outcomes.

Evidence boundary: these are controlled retrospective experiments using synthetic and oracle-informed candidates on real historical price series. They are not genuine out-of-sample model forecasts, live trading, audited investment performance, or evidence of guaranteed returns. External replication and prospective validation remain outstanding.

Limitations and appropriate use

Appropriate use requires held-out or walk-forward evaluation, leakage controls, realistic baselines, enough observations, stable data definitions, and application-specific outcome checks. For finance-related inputs, include realistic costs, position rules, liquidity assumptions, and risk constraints.

Known risks include sparse-event sensitivity, magnitude blindness in some candidate patterns, unstable regimes, selection bias, overlapping-window dependence, metric gaming, poor calibration, and distribution shift. Results from synthetic or oracle-informed candidates may not transfer to trained forecasting systems.

Neither metric should be used alone for investment, medical, safety-critical, credit, fraud, or other consequential decisions. Human review and domain-appropriate validation remain necessary.

FIS and CER FAQ

Does a high FIS mean a forecast will be profitable?

No. FIS is not an expected-return estimate or a probability of success. Profitability depends on data quality, costs, execution, risk, future conditions, and the decision rule.

Should FIS replace MSE or MASE?

No. FIS provides a complementary perspective. Error metrics reveal weaknesses that FIS may not capture, including important magnitude failures.

Why can FIS and CER rank models differently?

They provide distinct proprietary evaluation perspectives. A disagreement should trigger review of forecast behavior, conventional errors, validation stability, and application outcomes.

Are the experiments live trading results?

No. They are controlled retrospective model-selection experiments with synthetic and oracle-informed candidates. They do not represent deployable forecasts or investment returns.

Can scores be compared across unrelated datasets?

Not safely without evidence that the datasets, horizons, preprocessing, and validation settings are comparable. Within-study candidate rankings are the preferred interpretation.

Are the formulas public?

No. QuantSynth publishes intended use, interpretation, evidence, and limitations while retaining the mathematical specification and calculation procedure as proprietary intellectual property.

Proprietary methodology disclosure

FIS and CER use proprietary quantitative methodologies developed for QuantSynth. Their exact mathematical specifications, component calculations, transformations, aggregation procedures, weights, thresholds, exponents, calibration parameters, and implementation details are not publicly disclosed.

QuantSynth instead provides transparency about intended use, interpretation, evaluation requirements, empirical testing, known failure modes, and evidential limits. This supports informed assessment without enabling reconstruction of the protected scoring methods.

API status

Browser tools use internal JSON endpoints. QuantSynth does not currently offer a public programmatic API or API credentials. Internal browser endpoints are not a supported integration contract. Check with support for availability updates.