RisksManage, risk management engine by Fidel Kouessi

Live

Try the engine

Change the numbers: RisksManage recalculates everything instantly, in your browser. No data leaves your screen.

Estimated probability Odds offered Market margin Capital Sizing policy

Implied probability
Edge
Expected value per unit
Expected value at fair price
Full Kelly
Share committed
Recommended stake
Gain if successful
VaR 95%
CVaR 95%

Simulated capital with this policy: median , worst 5% of cases , probability of losing more than half .

Simplified demo. The full version works on your real positions, with historical VaR, replayed stress tests, sensitivities, line-by-line liquidity and every limit of your business.

An estimate is not a decision

The engine predicts nothing itself. It takes a probability and a real price, then passes them through a series of filters: each answers a precise question and hands its result to the next. For every rejected decision, you know where and why it was rejected.

01The fair price, stripped of the margin 02Value: the gap between two probabilities 03Calibration: learning to doubt your model 04Position size: Kelly, never in full 05Naming the loss: VaR and CVaR 06Monte Carlo: simulating to see the whole distribution 07Dependencies and concentration 08Live positions: an option to manage 09Proof: a ledger no one can rewrite 10Financial portfolios: drawdown and limits

Asset management company

What RisksManage measures in an asset management company

Every day, fund by fund, RisksManage calculates the measures required by the risk department, the custodian, the statutory auditor and the regulator, and raises an alert as soon as a limit comes close.

Market risk

Historical, parametric and Monte Carlo VaR, CVaR, at 95% and 99%, over one day and twenty days. Every VaR model is backtested: the Kupiec test checks that the number of observed breaches is consistent with the stated level.

{V a R}_{99 %}^{hist} = - Q_{1 %} (Δ V_{1}, \dots, Δ V_{T})

Q: quantile of value changes over T days of history

L R = - 2 \ln \frac{(1 - p)^{T - x} p^{x}}{(1 - \hat{p})^{T - x} {\hat{p}}^{x}},

x breaches observed over T days, p the expected rate

Stress tests

Historical crises replayed on today's portfolio, and hypothetical shocks on each factor: rates, equities, credit, currencies, commodities. The result is compared with the loss limits set by management.

Δ V \approx \sum_{k} \frac{\partial V}{\partial f_{k}} Δ f_{k} + \frac{1}{2} \sum_{k, l} \frac{\partial^{2} V}{\partial f_{k} \partial f_{l}} Δ f_{k} Δ f_{l}

f: risk factors, second-order expansion

Example of a diversified fund: two scenarios breach the −10% limit and trigger an alert.

Sensitivities

Modified duration and convexity of bonds, delta and vega of options, beta of equities: what the fund loses or gains for a given move in each market.

\frac{Δ P}{P} \approx - D_{mod} Δ y + \frac{1}{2} C (Δ y)^{2}

Liquidity

How many days does it take to sell each line without moving the market, and can the fund meet massive redemptions? The liquidity profile answers, horizon by horizon.

j_{i} = \frac{Q_{i}}{θ \cdot {A D V}_{i}}

Q: quantity held, ADV: average daily volume, θ: share of volume allowed

\frac{assets sold within h days}{stressed redemptions} \geq 1

38% of assets can be sold in one day, 62% in one week: stressed redemptions are covered three times.

Concentration and regulatory ratios

Exposures by issuer, sector, country and currency, the UCITS 5/10/40 rule, effective number of lines, leverage: every ratio is monitored continuously, with an alert before any breach.

w_{i} \leq 10 %,

the UCITS 5/10/40 rule

H H I = \sum_{i} w_{i}^{2},

N_eff: the number of truly independent lines

L = \frac{\sum_{j} ∣ E_{j} ∣}{N A V}

E: exposure of each instrument, NAV: net asset value

Risk-adjusted performance

Volatility, tracking error, Sharpe ratio, information ratio and drawdown: performance measured against the risk taken to achieve it, fund by fund and manager by manager.

T E = σ (R_{P} - R_{B}) \sqrt{252}

S = \frac{{\overset{ˉ}{R}}_{P} - R_{f}}{σ_{P}},

Counterparty and currency

Exposures by counterparty, collateral received and posted, net exposure by currency after hedging: the risks that do not show in market value.

The fair price, stripped of the margin

Odds, a price, a rate: every market price contains an implied probability. The engine reads it, then removes what the intermediary takes. Added together, the implied probabilities of an event always exceed 100%: this excess, the overround, is the margin.

π_{i} = \frac{1}{c_{i}}

O = \sum_{i} π_{i} - 1

p_{i}^{*} = \frac{π_{i}}{\sum_{j} π_{j}}

the normalised reference probability

A consequence many tools ignore: a position taken exactly at the fair price does not have zero expected value, it has a negative one, equal to the cost of the margin. The engine calculates this reference expected value and judges every position against it.

{E V}_{ref} = \frac{1}{1 + O} - 1 \approx - O

What it bringsKnowing what a position really costs before taking it, and comparing prices from different venues on a clean basis.

Three outcomes priced at 2.15, 3.50 and 3.30: 105.4% in total, i.e. 5.4 points of margin. The fair price brings the total back to 100%.

Value: the gap between two probabilities

The value of a position is measured by comparing the probability estimated by a model with the probability implied by the price. The gap is called the edge; the expected gain per unit committed follows directly.

edge = p - \frac{1}{c}

E V = p c - 1

A model that gives an outcome 46% when the price of 2.50 implies only 40% has 6 points of edge:

0.46 \times 2.50 - 1 = + 0.15

that is 15 cents of expected value per unit committed

Three verdicts come out of it: value when expected value and edge are truly positive, with no tolerance; fair when the price is within three points of the reference expected value; rejected in all other cases. A calculated price never replaces a real price: value is never measured against a price that does not exist.

What it bringsRanking hundreds of opportunities on a single comparable measure, and automatically rejecting those that are worth only their marketing.

Verdict	Condition
Value	expected value and edge strictly positive
Fair	expected value within 3 points of the reference
Rejected	all other cases

Three verdicts, one rule per verdict: the decision stays readable and open to challenge.

Calibration: learning to doubt your model

A model can often be right and still be poorly calibrated. Calibration does not measure how often answers are right, it measures how honest the percentages are: a well-calibrated model that announces 70% sees the event happen about seven times out of ten.

The engine applies a shrinkage factor specific to each domain, then refines it on real results, with Bayesian caution: nothing moves below thirty observations, and the weight of evidence only reaches its maximum at two hundred.

f = (1 - w) f_{0} + w \frac{\overset{ˉ}{r}}{\overset{ˉ}{p}},

r̄: observed hit rate, p̄: average announced probability

\tilde{p} = \min ⁣ (0.97, \max (0.01, f p)),

never certainty: the probability is bounded between 1% and 97%

What it bringsReducing model risk: an inflated probability mechanically leads to oversized positions. This is the layer that sets a serious tool apart from a percentage generator.

Reliability diagram: below the diagonal, the model promises more than it delivers. Calibration brings it back towards the diagonal.

Position size: Kelly, never in full

Finding a position with positive expected value solves only half the problem: an excellent position that is too large can ruin a capital base. The Kelly criterion gives the fraction that maximises long-term capital growth.

f^{*} = \frac{b p - q}{b},

G (f) = p \ln (1 + b f) + q \ln (1 - f)

expected growth per decision, maximal at f*

Full Kelly assumes a probability known with certainty and produces capital swings that few organisations can bear. The engine therefore calculates full, half and quarter Kelly, and recommends half Kelly by default, which keeps about three quarters of the growth for half the variance. Two caps apply: never more than 50% at full Kelly, never more than 25% as a recommendation.

What it bringsTurning an edge into an amount committed, reproducibly, instead of leaving position size to each person's temperament.

Probability 55%, odds 2.00: Kelly recommends 10%. At double, growth falls back to zero; beyond that, it turns negative.

Naming the loss: VaR and CVaR

Value at Risk at 95% is the loss not exceeded in 95% of cases. CVaR, or Expected Shortfall, goes further: it is the average loss in the worst 5% of cases, the ones that really matter.

{V a R}_{α} = \inf {ℓ : P (L \leq ℓ) \geq α}

{C V a R}_{α} = E [L ∣ L \geq {V a R}_{α}]

For a binary position, the calculation is exact and instant: as soon as the probability of losing exceeds 5%, the 95% VaR is the whole stake. For a portfolio, the engine combines the historical approach, simulation and the parametric approach:

{V a R}_{95 %} \approx 1.645 σ \sqrt{h} V

σ: volatility, h: horizon, V: exposed value

What it bringsShowing the quantified adverse scenario before any decision. It is the common language of risk between teams, management and regulators.

Simulated distribution of outcomes: VaR cuts off the worst 5% of cases, CVaR gives their average.

Monte Carlo: simulating to see the whole distribution

When outcomes combine, closed-form formulas are no longer enough. The engine simulates thousands of scenarios, each event drawn as a Bernoulli variable, and derives the real probability of winning, the average return, VaR, CVaR, and the best and worst outcomes.

P (all win) = \prod_{i = 1}^{k} p_{i},

five events at 65% each: less than one chance in eight

The same technique projects capital over time. Over 300 decisions at 55%, half Kelly and full Kelly have almost the same median but very different dispersion: in the worst 5% of cases, capital ends at 76 with half Kelly, against 27 with full Kelly.

What it bringsTesting a risk policy before applying it, and showing the dispersion of outcomes rather than a misleading average.

4,000 simulated paths: 50% and 90% bands. Dispersion, not the median, is what separates two risk policies.

Dependencies and concentration

Multiplying probabilities only makes sense if events are independent. Two positions on the same underlying, the same match or the same factor are not. The engine counts these collisions and flags any combination that is not independent.

σ_{P}^{2} = \sum_{i} w_{i}^{2} σ_{i}^{2} + 2 \sum_{i < j} w_{i} w_{j} ρ_{i j} σ_{i} σ_{j}

portfolio variance depends on correlations, not only on individual risks

Ignoring a positive correlation means underestimating the real risk at the exact moment everything moves in the same direction. Fine modelling of correlations across markets is still work in progress, and we prefer to say so.

What it bringsSpotting hidden concentrations and preventing apparent diversification from masking one and the same exposure.

Correlation matrix: A and B are on the same event, C and E share a common factor.

Live positions: an option to manage

Once the event has started, the question changes: hold, sell or hedge. A live position behaves like a binary option, paying a fixed amount if it finishes as a winner and nothing otherwise.

V_{t} = P_{final} \times G

G: the payout if successful

The engine borrows four layers from financial mathematics: the instantaneous probability read from the live price, a Poisson process for events still to come, a weighted sum over the possible number of jumps inspired by the Merton model, and volatility that rises with tension, in the manner of the Heston model.

N_{τ} \sim P o i s s o n (λ τ)

P_{final} = \sum_{n = 0}^{N} e^{- λ} \frac{λ^{n}}{n!} p_{n}

The rules stay readable: cash out above 85% when at most half the time remains, cut below 25%, consider a hedge after a fall of more than twenty points. The hedge stake that equalises both outcomes is:

h = \frac{S c_{1}}{c_{2}}

S: initial stake at odds c₁, c₂: odds of the opposite outcome

What it bringsA quantified and reasoned take-profit and stop-loss, where until now there was only a button and an amount offered by the counterparty.

Decision map: a path that enters the green zone is locked in; a sharp fall triggers a hedge review.

Proof: a ledger no one can rewrite

A risk engine is only worth something if what it produces can be verified. Every decision is sealed when it is taken, then published in a ledger where each entry contains the fingerprint of the previous one.

h_{k} = S H A 256 (h_{k - 1} ∥ m_{k} ∥ s_{k})

m: content, s: salt; any change breaks the chain

The outcome of a decision is a poor judge of its quality: it takes hundreds of observations for luck to fade. Closing Line Value gives a much faster signal, by comparing the price obtained with the fair closing price.

C L V = c_{taken} \times p_{close} - 1

What it bringsA timestamped audit trail that a third party can verify: useful for compliance, internal control and client trust.

Each entry carries the SHA-256 fingerprint of the previous one: anyone can recompute the chain.

Financial portfolios: drawdown and limits

Applied to an asset portfolio, the engine tracks value, volatility, VaR and loss from the peak, and continuously checks compliance with set limits: exposure by pocket, by asset and by strategy.

{D D}_{t} = 1 - \frac{V_{t}}{\max_{s \leq t} V_{s}}

A guardian watches positions and can suspend everything. Humans always stay in charge: stopping automation, selling or liquidating everything remains possible at any moment.

What it bringsFraming automated or discretionary management with quantified safeguards, and documenting every breach.

Portfolio value, all-time high and drawdown: the measure of the pain actually endured.

For each business, its risks

The same engine, the same standards of proof, applied to the risks specific to each business.

Business	What RisksManage measures	Key measures
Asset managers, family offices	Market, liquidity, concentration, leverage, regulatory ratios	VaR, CVaR, stress tests, liquidity profile, 5/10/40, tracking error
Banks	Balance sheet interest rate and liquidity, market	Interest rate gap, sensitivity of economic value and net interest income, LCR, NSFR, VaR
Insurers	Event risks, investments	Calibrated probabilities, Monte Carlo of claims, VaR of investments
Betting operators	Exposure per event, margin, winning customers	Overround, net exposure per outcome, customer CLV
Prediction markets, market makers	Prices, inventory, open positions	Fair price, Kelly size, value of open contracts
Brokers, investment fintechs	Client portfolios, margin calls	VaR per client, drawdown, automatic safeguards
Marketplaces, delivery	Escrow, disputes, fraud	Money held until delivery, dispute rate, exposure per seller
Corporate treasuries	Currencies, rates, commodities	Potential loss on exposures, size of hedges

Example: a bank's asset and liability management

The economic value of equity and the liquidity coverage ratio, recalculated for each regulatory interest rate scenario.

Δ E V E = \sum_{t} {C F}_{t} (e^{- (y_{t} + Δ y_{t}) t} - e^{- y_{t} t})

L C R = \frac{H Q L A}{30-day net outflows} \geq 100 %

Why adopt it

What the company gains

A single framework

Market, liquidity, concentration, performance and decisions in one tool, with the same definitions for everyone.

Figures that explain themselves

Every measure comes with all its intermediate values. No black box in front of the risk committee or the regulator.

A tamper-proof audit trail

Every decision is timestamped and sealed in a chained ledger: ready for the custodian, the statutory auditor and inspections.

Continuous alerts

Limits are monitored at all times, not just at month end, and every breach is documented.

Quick to integrate

An API, a core with no external dependencies, no heavy IT project and no big-vendor licence.

Proven on extreme markets

Live sport, cryptocurrencies, prediction markets: if the engine holds there, it holds on a bond fund.

Designed by a practitioner

Built by a specialist in market risk, asset and liability management and fund audit, who knows the controls from the inside and has already helped design, test and sell a risk management tool to asset management companies.

Open to African markets

Multi-currency from the outset, including the continent's exchanges and currencies.

Already in service

Risk coverage across our projects

Before being offered to companies, the engine protects Fidel Kouessi's projects, each on its own risks.

PROLIFICK

Fair price and margin, calibration by sport, fractional Kelly, four staking profiles, 95% VaR, session limit, Monte Carlo of accumulators, live cash-out and hedging.

Robot PROLIFICK

Automated positions on sports events, limits per bet and per day, CLV tracking; in testing with simulated money.

IFEDUN

Limits per pocket and per strategy, drawdown monitoring, downside hedging, live account guardian, full stop in one move, risk reports.

Certipick

Odds captured by the server, CLV published, drawdown and standard deviation displayed, track record sealed in a chained ledger.

Probdesk

Prices stripped of the margin, gaps between Polymarket, Kalshi and bookmakers, best buy fees included.

AHI and Zeyini

Money held in escrow until safe delivery, handover code, verified sellers and providers: counterparty and fraud risk under control.

How it integrates

A transparent API

Every call returns all intermediate values: raw and calibrated probability, implied probability, edge, expected value, Kelly fraction, VaR, score. A recommendation explains itself, it does not just assert.

Risk reports

Summaries by portfolio, by strategy and by period: exposures, potential losses, limit compliance, sealed history.

A core tested in isolation

The engine depends on no external library: every function can be tested on its own, with no database or network. A tool that claims to bring rigour must first be rigorous itself.

What the engine does not promise: it does not remove chance and does not turn any position into a safe investment. It improves the quality of the reasoning before the decision, and makes that reasoning verifiable.