Strategies
Buffered momentum strategy
Selects securities with persistent positive momentum and reduces unnecessary portfolio churn.
The holding buffer helps avoid portfolio changes after small ranking moves.
Loading the section
Clear explanations of strategies, metrics, sources, the agent layer, data and demo execution, without internal engineering language.
Selects securities with persistent positive momentum and reduces unnecessary portfolio churn.
The holding buffer helps avoid portfolio changes after small ranking moves.
Reduces market exposure when broad market volatility becomes too high.
It is a protective overlay over the base strategy: in an unfavorable regime, part of the risk moves to cash.
Uses the market trend and favors securities with strong momentum relative to risk.
If the market regime is weak, the strategy can reduce risk exposure instead of mechanically buying recent winners.
Buys securities only when their own momentum is positive and the market regime is favorable.
This scenario filters out securities that only look strong relative to an even weaker market.
Checks whether the broad market is suitable for taking market risk.
The filter reduces the chance of entering a portfolio during broad market weakness.
Compares each security's momentum with its volatility, avoiding purely noisy moves.
The security weight depends not only on return, but also on movement stability.
Buys securities whose own performance over the selected window is positive.
If positive momentum is absent, the scenario can move into a defensive mode.
Selects securities that outperformed peers over the previous observation window.
The simple version is useful as a baseline for stricter strategies.
Buys securities with positive own momentum and sizes them by risk.
Less volatile securities receive higher weight, all else equal.
Combines trading value, price momentum and volatility to rank securities.
The scenario is useful when selecting from a broad MOEX catalog without manual ticker picking.
A state where the selected scenario considers market risk acceptable.
In this mode, the service may propose a securities portfolio instead of cash.
Allocates capital equally across selected securities.
Used as a clear baseline for comparison with agent-selected scenarios.
Keeps the initial portfolio almost unchanged.
The scenario is used to compare active selection with a simple long-term approach.
A defensive mode where the system avoids taking market risk.
This choice is expected when the signal is weak, data is incomplete or risk looks excessive.
The system abstains when data or signal quality is insufficient.
Abstaining from a position is a valid outcome, not a service error.
Builds a portfolio from the current market state under the user's risk profile.
The scenario uses fresh quotes, the security catalog, constraints and saved preferences.
Opens a previously calculated strategy selection with charts, metrics and explanation.
This supports reproducible review of long windows and prevents accidental selection changes.
A strategy calculation over a long historical window to test result stability.
Such runs help compare agent choices with baselines without manual tuning.
Worst peak-to-trough drop in portfolio value.
A key indicator of how painful the strategy path could have been.
Estimated loss level not expected to be exceeded in 95% of ordinary observations.
This is a statistical estimate, not a promise of maximum loss.
Average loss in the worst 5% of observations.
The metric shows the severity of adverse days, not only a loss threshold.
Portfolio sensitivity to market-index moves.
High beta means the result depends more strongly on the broad market.
How actively the strategy reallocates capital between securities.
High turnover increases sensitivity to commissions and slippage.
Commissions and slippage accounted for in the simulation.
Without costs, active strategies can look better than they would in realistic execution.
The market state that determines whether to take risk or move to a defensive scenario.
The risk that a security is hard to buy or sell without meaningful price impact.
Decision error risk caused by stale, incomplete or inconsistent data.
The risk that an explanation sounds convincing but is weakly tied to evidence and calculation.
Annualized capital growth rate over the period.
The metric is convenient for comparing strategies with different history lengths.
How much capital increased or decreased over the whole period.
Shows the overall result, but not the path or drawdown depth.
Return per unit of total volatility. Higher is better for risk-adjusted performance.
Useful for comparing strategies, but does not separate ordinary volatility from downside moves.
Return per unit of downside volatility, focusing on negative moves.
Better suited when downside control matters more than total volatility.
Annualized return divided by maximum drawdown.
Helps assess whether the strategy's return compensates for its capital drops.
Return beyond what is explained by market movement.
Positive alpha indicates strategy value added relative to the market factor.
How consistently the strategy outperforms a benchmark relative to tracking error.
The metric is useful when the strategy is judged relative to an index.
Volatility of the return difference between strategy and benchmark.
Higher values mean the strategy deviates more from index behavior.
A chart of portfolio value over time.
The curve shows not only the final result, but also the path to it.
Shows capital decline from previous highs.
Useful for assessing the depth and duration of adverse periods.
A market reference used to compare a strategy or portfolio.
Without a benchmark, it is hard to judge whether the strategy added value versus the market.
Shows which securities and decisions affected portfolio performance.
Ranking quality: how high the model places genuinely strong securities.
The metric fits security selection tasks, not only price forecasting.
Correlation between predicted security ranks and future returns.
High RankIC means the model ranks securities better by future performance.
Mean absolute forecast error.
Shows the average error size in the original target scale.
Root mean squared error, penalizing large misses more strongly.
The metric is especially sensitive to rare but large forecast errors.
The historical period used to fit model or strategy parameters.
A period used to validate tuning without using future data.
Testing on a period not used for parameter selection.
A list of trained models with statuses, metrics and data used.
A fact, news item or document used to support the explanation.
Evidence does not replace calculation, but helps explain decision context.
A direct link between a claim and a document, news item, fact or event.
A compact semantic part of text or an event retrievable through Qdrant.
A search structure for quickly finding semantically related documents and events.
The service uses the index for news, documents, issuer events and decision context.
An explanation where material claims are tied to retrieved evidence.
A decision uses only information available at the decision date.
This prevents the service from looking into future reports, news and quotes.
The documents and events retrieved before the agent forms an explanation.
A company summary with quotes, sector, news, reports, dividends and retrieved events.
A check of completeness, publication date and relevance of evidence for the decision.
An allowed service function with typed input, output, timeout and audit.
The service layer that calculates weights, metrics, constraints and risk checks by fixed rules.
The language model that parses the request, calls allowed tools and explains results.
Portfolio weights and risk limits remain in the calculation layer.
A Qwen model deployed next to the service for request parsing and explanation drafting.
Local mode supports the agent demonstration without an external LLM provider.
Saved preferences: risk profile, horizon, watchlist and previous recommendations.
Checks preventing the agent from violating constraints, leaking secrets or bypassing risk checks.
A mechanism that prevents external text from changing the agent's operating rules.
A background recommendation calculation with status, start time, completion and saved result.
The current agent job state: data preparation, calculation, explanation, saving or completion.
Step-by-step log: request, data, features, portfolio, risks and explanation.
A calculation result that appears in history and can be reopened.
A record of user and service actions for reproducibility, incident analysis and control demonstration.
How close the market state is to the latest available trading date.
How many securities in the selected universe have quotes and parameters for calculation.
Data older than the accepted freshness window, so the recommendation should be recalculated carefully.
The source returned only part of the data. The service must show what is missing.
A MOEX security list with ticker, company name and interface search.
A security price chart with open, high, low and close values over time.
A visual summary of securities: size, price change, liquidity and sector.
Saved quotes and reference data for a specific moment in time.
Issuer financial reporting used as context after its publication date.
A list of expected and historical dividend events by security.
A shared calculation of trend, momentum, liquidity and risk quality by security.
A saved set of numeric features prepared strictly as of the calculation date.
A readiness check for key pages, API routes, data and the demonstration scenario.
A fixed set of data, reports and instructions for a reproducible service demonstration.
A non-live account showing simulated execution and operation history.
The service does not send broker orders. It shows a proposal, risks and demo execution.
A virtual portfolio showing the decision impact on capital and positions.
A calculated list of buys and sells that can be reviewed before execution.
A log of virtual buys, sells, capital changes and execution statuses.
The date for which the signal is calculated and source data is available.
The date to which simulated execution of the decision is mapped.
The gap between expected and actual execution price that should be included in realistic simulation.
A recommendation calculation in the product interface with data preparation, checks and result persistence.
A result with artifacts, state codes and agent trace that can be opened and compared.
No. The service supports trading decisions: it calculates a proposal, shows risk, explains the logic and provides demo execution without real broker orders.
It gathers the risk profile, watchlist, quotes, sources and features, runs allowed calculations, saves the recommendation and prepares the explanation.
When the job has a final status, the recommendation appears in history, and the explanation plus agent trace are saved and available to open.
Weights, constraints and risk checks are calculated by the deterministic layer. The LLM parses the request, calls allowed tools and explains the result.
On a decision date, only quotes, news, reports and facts already available at that time are used. This prevents the result from looking into the future.
This happens when market regime, risk, liquidity or data quality make entering securities worse than the defensive scenario.
Quotes show price movement, while sources add context: news, corporate facts, reports, documents and events that can support the explanation.
Material claims should be tied to retrieved sources. The analytics view can open the used context, dates, fragments and decision linkage.
The service shows data freshness and coverage. If data is missing, the recommendation is less reliable and must clearly explain what is missing.
It quickly shows which securities move stronger, where liquidity is higher and how changes are distributed by sector. It is an overview, not a standalone recommendation.
Analytics reviews decisions, sources, charts and security contribution. Models and Metrics shows the quality of strategies, models, runs and comparisons.
The demo account shows simulated execution without real orders: operation plan, virtual capital changes, history and portfolio impact.
No. External texts are treated as data sources, not commands. Protective checks prevent them from changing rules, bypassing risk control or accessing secrets.