Six portfolio presets. Seven scoring dimensions. Two AI systems. Most robo-advisers won't tell you any of this. We document it because the quality of our process is a better sales argument than a vague promise.
A preset is not a fixed portfolio — it is a weighting profile that tells the scoring engine which dimensions matter most for your investment objective. You choose the preset; the engine builds and maintains the portfolio within your configured limits.
Prioritises Volatility Profile and Fundamental Quality. Targets lower-beta equities, dividend-paying stocks, and assets with stable earnings histories. Suitable for investors with a shorter time horizon or lower risk tolerance.
Equal-weighted scoring across all seven dimensions. Broadly diversified across sectors and geographies. Designed to approximate long-term market exposure without strong sector or factor tilts.
Upweights Portfolio Correlation to maximise geographic and sector diversification. Actively reduces concentration risk by penalising high-correlation additions to your existing holdings.
Overweights Momentum Score and Fundamental Quality. Targets equities with strong recent price performance and improving earnings fundamentals. Higher expected volatility in exchange for higher expected return.
Sector-weighted preset that upweights technology, semiconductors, software, and AI-adjacent industries. Applies standard scoring dimensions within that universe. Higher concentration risk by design.
Applies an ESG screen as a pre-filter before scoring. Assets that fail minimum ESG thresholds are excluded from the universe entirely. Within the screened universe, standard scoring applies with a slight upweight on Fundamental Quality.
Every asset in your universe is scored on each dimension, normalised to a 0–100 scale, and weighted by your preset's priority profile. The weighted sum is the composite score used for portfolio decisions.
Measures recent price performance relative to the asset's own history and its sector peers. Inputs: 20-day, 60-day, and 120-day price return; RSI(14); distance from 52-week high/low. Source: FMP price data.
Evaluates earnings stability, balance sheet health, and valuation. Inputs: trailing P/E ratio, EPS growth (year-over-year), revenue growth, gross margin trend, debt-to-equity. Source: FMP fundamentals.
FinBERT-scored sentiment across recent Benzinga articles and GDELT tone signals for the asset's ticker and sector. Inputs: 7-day rolling sentiment average, volume-weighted by article recency. Negative outliers are penalised non-linearly.
Measures how well the asset fits your preset's sector weighting profile. An asset in an underweight sector scores higher than an equivalent asset in an already overweight sector. This dimension prevents sector concentration drift.
Measures historical volatility relative to the broad market. Inputs: 30-day realised volatility, beta vs S&P 500, maximum drawdown over 90 days. Conservative presets apply a stronger penalty for high volatility scores.
Ensures the system does not build positions in assets that would be costly or slow to exit. Inputs: 30-day average daily volume, bid-ask spread estimate, market capitalisation. Assets below a minimum liquidity threshold are excluded from the universe.
Measures how the candidate asset correlates with your existing holdings over a 90-day window. Assets that would reduce overall portfolio correlation score higher — effective diversification is rewarded. Assets that duplicate existing exposure score lower.
The news sentiment dimension is the most differentiated part of the scoring model. Standard quantitative factors — momentum, fundamentals, volatility — are publicly available and well-understood. The ability to process and act on real-time news context at scale is not.
Coincruze uses two AI systems in sequence for this: FinBERT classifies incoming articles into positive, negative, or neutral sentiment with a confidence score. Claude Haiku is then used downstream — not to re-classify sentiment, but to translate the system's final scoring output into plain-English decision rationales for your activity log.
The separation is deliberate. FinBERT is a smaller, faster, purpose-built financial NLP model — well-suited to the high-volume, low-latency task of article classification. Claude Haiku handles the lower-frequency, higher-quality language task of writing explanations a non-expert investor can actually understand.
Benzinga articles and GDELT records are pulled on a scheduled cycle, tagged to tickers, and queued for classification.
Each article is passed to FinBERT — a BERT-base model fine-tuned on financial phraseology (earnings reports, analyst notes, financial news). Output: [positive | neutral | negative] + confidence 0–1.
Article-level scores are aggregated into a 7-day rolling ticker sentiment score, weighted by publication recency and source credibility tier.
The rolling score is normalised against the asset's own 90-day sentiment baseline to distinguish between 'bad news' and 'worse than usual news'.
Once the scoring pipeline produces a trade candidate, Claude Haiku writes the rationale entry: which dimensions triggered the action, what the threshold was, and why the trade was proposed in plain English.
Scoring is continuous. Actions are event-driven. The system does not trade on every scoring cycle — it acts when a specific condition is met.
Assets in your universe are rescored on each data refresh cycle. Scores update as new prices, fundamentals, and news arrive. No action is taken at this stage.
After each scoring cycle, the system compares your current portfolio allocation to the target weights implied by current scores. If no asset exceeds your drift threshold, nothing happens.
If an asset's allocation drifts beyond your configured threshold (e.g., a position is 4% above target in a ±3% band), a rebalance candidate is generated.
Before a candidate becomes a trade, it is validated against your position limits, cash reserve setting, and single-trade size constraint. Any candidate that would breach a limit is rejected.
Validated candidates are submitted to IBKR as market or limit orders per your order type preference. Coincruze waits for IBKR's fill confirmation before marking the action complete.
Claude Haiku writes a plain-English rationale for the action. The entry is added to your decision log with the composite score, the specific dimension scores, and the threshold that triggered rebalancing.
The scoring model identifies signals correlated with historical performance. It cannot predict future prices.
No scoring output can cause the system to exceed your configured position limits, allocation ranges, or single-trade caps.
The model does not adapt its weights based on the performance of your specific account. Updates to the model apply globally after internal testing.
If FMP, Benzinga, or GDELT data is stale beyond a defined threshold, the scoring cycle is skipped rather than acting on incomplete information.
Every triggered action, rejected candidate, and score update is logged and visible in your decision history.
Drift thresholds, position limits, cash reserves, and order preferences are hard constraints. The model operates within them, always.
Most automated investment platforms describe their methodology in one sentence: "we use a proprietary algorithm." That tells you nothing. It makes verification impossible and creates the perception — often accurate — that the vagueness is intentional.
Coincruze's target users are young investors who are sceptical of financial services by default. They have grown up reading about hidden fees, conflicted advice, and systems that reward the house more than the investor.
Publishing the methodology in detail is how we earn credibility with that audience — not by asking for their trust, but by making it possible to verify the process before handing anything over.
Six presets, seven scoring dimensions, and a decision log that explains every move in plain English.
Investing involves risk. Scoring models are based on historical data and do not guarantee future performance. The methodology described reflects Coincruze's current implementation and may be updated. All portfolio actions are subject to your configured limits and thresholds.