Inside the Trade Execution Proposal Engine

By now you can read a market. The harder question is what to do about it — and, before that, how much of what you are reading is real.

Every chapter of this institute has circled the same throughline: the discipline of judging when a number is genuine evidence that an asset will move, rather than noise wearing a convincing costume. The Trade Execution Proposal engine — the TEP that sits at the heart of SVQ — is where that judgment finally has to commit to a number you can lose money on. So let us walk into the engine and see how it turns a pile of signals into a single sized order, and why the math it uses is exactly the math of betting under uncertainty.

The engine has two jobs, and they happen in sequence. First it has to compose — to take many separate signals, each noisy and each disagreeing slightly with the others, and fuse them into one honest probability that the trade works. Then it has to size — to decide what fraction of your capital that probability justifies committing. Composition answers how sure are we. Sizing answers so how much. Get the first wrong and the second is poison, no matter how elegant the formula.

Start with composition. You have a signal stack: a momentum reading, a mean-reversion reading, an order-flow imbalance, a regime flag from SVNexus, perhaps a microstructure cue from the order book you studied last chapter. Each one, on its own, is a weak and biased forecaster. The naive instinct is to average their opinions, or to count how many point the same way and go with the majority. Both are wrong, and they are wrong in the same way: they treat a signal's vote as if it were its evidence.

The right object to reason about is not the signal's direction but the log-odds it contributes. Think of it this way. You begin with some prior belief that the asset moves up — say a base rate near one-half if you have no view. Each signal then nudges that belief, and the clean, additive currency for a nudge is the logarithm of a likelihood ratio: how much more often this signal reads the way it just read when price is about to rise than when it is about to fall. A signal that is barely better than a coin contributes a nudge near zero. A signal that is genuinely informative contributes a large nudge. And — this is the heart of it — these nudges add. You sum the log-odds contributions, weighted by how much you trust each signal, and convert the total back into a probability with the logistic function. This is just Bayesian updating written in its additive form, and it is the same log-space arithmetic you met in your very first foundations chapter, now doing real work.

Two things fall out of this that the TEP cares about deeply. First, a signal that carries no edge contributes a log-odds near zero, so it cannot move the composed probability — the stack is self-protecting against dead weight, if each signal's weight is honestly calibrated. Second, correlated signals must not be allowed to each shout the full nudge, because they are largely repeating the same evidence. Two momentum signals that are ninety percent the same thing are not two independent witnesses; counting them as two is how a stack convinces itself it is certain when it has heard a single fact twice. Composition done right is mostly the discipline of not double-counting.

The output of all this is a single number, p — the engine's calibrated probability that the trade pays off. Calibrated is the operative word. If the TEP says zero-point-six on a thousand occasions, the trade had better work about six hundred times. A probability that is not calibrated is not evidence; it is decoration.

Now the second job: sizing. Suppose the engine has its probability p, and the trade pays b dollars per dollar risked when it wins — that is your odds, your reward-to-risk. Let q be one minus p, the chance you lose. How much of your capital should you stake? The answer that maximizes the long-run growth rate of your wealth is the Kelly fraction:

f-star equals p times b, minus q, all divided by b.

Read it slowly, because every term earns its place. The numerator, p·b − q, is your edge: your probability-weighted winnings minus your probability-weighted losses, the expected profit per dollar of the bet. If that numerator is zero or negative, the trade has no edge and Kelly tells you to stake nothing — f-star is zero or it turns the bet down outright. Dividing by b scales that edge by the odds: the better the payoff per unit risk, the larger the fraction the same edge can support. So Kelly is simply edge over odds. It bets more when you are surer and when the payoff is richer, and it bets nothing the instant your edge evaporates.

Here is the discipline. Kelly is the theoretical maximum growth-optimal size, and it is brutally unforgiving of an overstated p. Because the numerator depends linearly on p but your survival depends on not blowing up, a probability that is even slightly too optimistic makes full Kelly bet too large, and the variance of full Kelly is savage — drawdowns that are correct in theory and intolerable in practice. This is why the TEP, like every serious shop, sizes at a fraction of Kelly — a half or a quarter — trading a little long-run growth for a large reduction in the gut-wrenching swings. And it is why, when composition cannot produce a p whose edge clears a threshold, the engine proposes no trade at all. A flat numerator is the engine telling you, honestly, that what you are looking at is not yet evidence.