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.
Product glossary
Glossary and FAQ
Clear explanations of strategies, metrics, sources, the agent layer, data and demo execution, without internal engineering language.
Search
93 terms · 13 questions
Terms
Terms and explanations
Strategies
Strategies
18
Strategies
Volatility-control strategy
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.
Strategies
Trend-filtered risk-adjusted momentum
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.
Strategies
Trend-filtered absolute momentum
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.
Strategies
Market-trend filter
Checks whether the broad market is suitable for taking market risk.
The filter reduces the chance of entering a portfolio during broad market weakness.
Strategies
Risk-adjusted momentum
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.
Strategies
Absolute momentum strategy
Buys securities whose own performance over the selected window is positive.
If positive momentum is absent, the scenario can move into a defensive mode.
Strategies
Momentum strategy
Selects securities that outperformed peers over the previous observation window.
The simple version is useful as a baseline for stricter strategies.
Strategies
Time-series momentum with inverse volatility
Buys securities with positive own momentum and sizes them by risk.
Less volatile securities receive higher weight, all else equal.
Strategies
Liquidity, momentum and risk ranking
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.
Selection modes
Risk-on mode
A state where the selected scenario considers market risk acceptable.
In this mode, the service may propose a securities portfolio instead of cash.
Baselines
Equal-weight portfolio
Allocates capital equally across selected securities.
Used as a clear baseline for comparison with agent-selected scenarios.
Baselines
Buy and hold equally
Keeps the initial portfolio almost unchanged.
The scenario is used to compare active selection with a simple long-term approach.
Defensive modes
Cash / no position
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.
Defensive modes
No entry
The system abstains when data or signal quality is insufficient.
Abstaining from a position is a valid outcome, not a service error.
User scenarios
Portfolio on current MOEX coverage
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.
Saved scenarios
Saved strategy-selection run
Opens a previously calculated strategy selection with charts, metrics and explanation.
This supports reproducible review of long windows and prevents accidental selection changes.
Saved scenarios
Long-window strategy run
A strategy calculation over a long historical window to test result stability.
Such runs help compare agent choices with baselines without manual tuning.
Risk
Risk
10
Risk
Maximum drawdown
Worst peak-to-trough drop in portfolio value.
A key indicator of how painful the strategy path could have been.
Risk
VaR 95%
Estimated loss level not expected to be exceeded in 95% of ordinary observations.
This is a statistical estimate, not a promise of maximum loss.
Risk
CVaR 95%
Average loss in the worst 5% of observations.
The metric shows the severity of adverse days, not only a loss threshold.
Risk
Market beta
Portfolio sensitivity to market-index moves.
High beta means the result depends more strongly on the broad market.
Implementability
Portfolio turnover
How actively the strategy reallocates capital between securities.
High turnover increases sensitivity to commissions and slippage.
Implementability
Costs
Commissions and slippage accounted for in the simulation.
Without costs, active strategies can look better than they would in realistic execution.
Risk
Market regime
The market state that determines whether to take risk or move to a defensive scenario.
Risk
Liquidity risk
The risk that a security is hard to buy or sell without meaningful price impact.
Risk
Data risk
Decision error risk caused by stale, incomplete or inconsistent data.
Risk
Explanation risk
The risk that an explanation sounds convincing but is weakly tied to evidence and calculation.
Return
Return
12
Return
CAGR
Annualized capital growth rate over the period.
The metric is convenient for comparing strategies with different history lengths.
Return
Total return
How much capital increased or decreased over the whole period.
Shows the overall result, but not the path or drawdown depth.
Risk-adjusted performance
Sharpe ratio
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.
Risk-adjusted performance
Sortino ratio
Return per unit of downside volatility, focusing on negative moves.
Better suited when downside control matters more than total volatility.
Risk-adjusted performance
Calmar ratio
Annualized return divided by maximum drawdown.
Helps assess whether the strategy's return compensates for its capital drops.
Benchmark comparison
Alpha
Return beyond what is explained by market movement.
Positive alpha indicates strategy value added relative to the market factor.
Benchmark comparison
Information ratio
How consistently the strategy outperforms a benchmark relative to tracking error.
The metric is useful when the strategy is judged relative to an index.
Benchmark comparison
Tracking error
Volatility of the return difference between strategy and benchmark.
Higher values mean the strategy deviates more from index behavior.
Return
Equity curve
A chart of portfolio value over time.
The curve shows not only the final result, but also the path to it.
Risk-adjusted performance
Drawdown chart
Shows capital decline from previous highs.
Useful for assessing the depth and duration of adverse periods.
Benchmark comparison
Benchmark
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.
Return
Performance attribution
Shows which securities and decisions affected portfolio performance.
Forecasting and ranking
Forecasting and ranking
8
Forecasting and ranking
NDCG@k
Ranking quality: how high the model places genuinely strong securities.
The metric fits security selection tasks, not only price forecasting.
Forecasting and ranking
RankIC
Correlation between predicted security ranks and future returns.
High RankIC means the model ranks securities better by future performance.
Forecasting and ranking
MAE
Mean absolute forecast error.
Shows the average error size in the original target scale.
Forecasting and ranking
RMSE
Root mean squared error, penalizing large misses more strongly.
The metric is especially sensitive to rare but large forecast errors.
ML/DL
Training window
The historical period used to fit model or strategy parameters.
ML/DL
Validation window
A period used to validate tuning without using future data.
ML/DL
Out-of-sample check
Testing on a period not used for parameter selection.
ML/DL
Model registry
A list of trained models with statuses, metrics and data used.
RAG and evidence
RAG and evidence
9
Evidence
Evidence
A fact, news item or document used to support the explanation.
Evidence does not replace calculation, but helps explain decision context.
Evidence
Source citation
A direct link between a claim and a document, news item, fact or event.
RAG
Chunk
A compact semantic part of text or an event retrievable through Qdrant.
RAG
Vector index
A search structure for quickly finding semantically related documents and events.
The service uses the index for news, documents, issuer events and decision context.
Evidence
Grounded explanation
An explanation where material claims are tied to retrieved evidence.
Evidence
Point-in-time approach
A decision uses only information available at the decision date.
This prevents the service from looking into future reports, news and quotes.
RAG
Retrieved context
The documents and events retrieved before the agent forms an explanation.
Evidence
Issuer dossier
A company summary with quotes, sector, news, reports, dividends and retrieved events.
Evidence
Source quality
A check of completeness, publication date and relevance of evidence for the decision.
Agent
Agent
12
Agent
Agent tool
An allowed service function with typed input, output, timeout and audit.
Agent
Deterministic calculation layer
The service layer that calculates weights, metrics, constraints and risk checks by fixed rules.
Agent
LLM supervisor
The language model that parses the request, calls allowed tools and explains results.
Portfolio weights and risk limits remain in the calculation layer.
Agent
Local language model
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.
Agent
Memory
Saved preferences: risk profile, horizon, watchlist and previous recommendations.
Security
Guardrails
Checks preventing the agent from violating constraints, leaking secrets or bypassing risk checks.
Security
Prompt-injection protection
A mechanism that prevents external text from changing the agent's operating rules.
Operations
Agent job
A background recommendation calculation with status, start time, completion and saved result.
Operations
Progress state
The current agent job state: data preparation, calculation, explanation, saving or completion.
Operations
Agent trace
Step-by-step log: request, data, features, portfolio, risks and explanation.
Operations
Saved recommendation
A calculation result that appears in history and can be reopened.
Operations
Audit log
A record of user and service actions for reproducibility, incident analysis and control demonstration.
Data
Data
14
Data
Data freshness
How close the market state is to the latest available trading date.
Data
Coverage
How many securities in the selected universe have quotes and parameters for calculation.
Data
Stale market state
Data older than the accepted freshness window, so the recommendation should be recalculated carefully.
Data
Partial data
The source returned only part of the data. The service must show what is missing.
Data
Security catalog
A MOEX security list with ticker, company name and interface search.
Market
Candlestick chart
A security price chart with open, high, low and close values over time.
Market
Market map
A visual summary of securities: size, price change, liquidity and sector.
Market
Market state
Saved quotes and reference data for a specific moment in time.
Data
Company report
Issuer financial reporting used as context after its publication date.
Data
Dividend calendar
A list of expected and historical dividend events by security.
Data
Technical signal state
A shared calculation of trend, momentum, liquidity and risk quality by security.
Data
Feature state
A saved set of numeric features prepared strictly as of the calculation date.
Readiness
Service check
A readiness check for key pages, API routes, data and the demonstration scenario.
Readiness
Demo package
A fixed set of data, reports and instructions for a reproducible service demonstration.
Demo execution
Demo execution
10
Demo execution
Paper account
A non-live account showing simulated execution and operation history.
Security
Real trading disabled
The service does not send broker orders. It shows a proposal, risks and demo execution.
Demo execution
Paper portfolio
A virtual portfolio showing the decision impact on capital and positions.
Demo execution
Order plan
A calculated list of buys and sells that can be reviewed before execution.
Demo execution
Operation history
A log of virtual buys, sells, capital changes and execution statuses.
Demo execution
Decision date
The date for which the signal is calculated and source data is available.
Demo execution
Execution date
The date to which simulated execution of the decision is mapped.
Implementability
Slippage
The gap between expected and actual execution price that should be included in realistic simulation.
Operations
Interactive calculation
A recommendation calculation in the product interface with data preparation, checks and result persistence.
Operations
Saved calculation
A result with artifacts, state codes and agent trace that can be opened and compared.
FAQ
FAQ
Is this a trading robot?
No. The service supports trading decisions: it calculates a proposal, shows risk, explains the logic and provides demo execution without real broker orders.
What does the agent do during a calculation?
It gathers the risk profile, watchlist, quotes, sources and features, runs allowed calculations, saves the recommendation and prepares the explanation.
When is a calculation complete?
When the job has a final status, the recommendation appears in history, and the explanation plus agent trace are saved and available to open.
Why does the LLM not calculate portfolio weights directly?
Weights, constraints and risk checks are calculated by the deterministic layer. The LLM parses the request, calls allowed tools and explains the result.
What does point-in-time mean?
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.
Why can the service choose cash / no position?
This happens when market regime, risk, liquidity or data quality make entering securities worse than the defensive scenario.
Why use sources if quotes already exist?
Quotes show price movement, while sources add context: news, corporate facts, reports, documents and events that can support the explanation.
How can I tell that the explanation is not fabricated?
Material claims should be tied to retrieved sources. The analytics view can open the used context, dates, fragments and decision linkage.
What happens with stale or partial data?
The service shows data freshness and coverage. If data is missing, the recommendation is less reliable and must clearly explain what is missing.
What is the market map for?
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.
How is Analytics different from Models and Metrics?
Analytics reviews decisions, sources, charts and security contribution. Models and Metrics shows the quality of strategies, models, runs and comparisons.
What does the demo account show?
The demo account shows simulated execution without real orders: operation plan, virtual capital changes, history and portfolio impact.
Can a news text change the agent rules?
No. External texts are treated as data sources, not commands. Protective checks prevent them from changing rules, bypassing risk control or accessing secrets.