What makes an uncertain promise worth something today? Is a price a forecast of the future, a measure of collective optimism, or an expression of how much we fear bad outcomes? Financial markets seem to bundle all of these forces into a single number. The fundamental theorem of asset pricing shows that, beneath that complexity, prices must obey a remarkably simple logic: consistent prices leave no room for a free lunch.

The setting. We'll illustrate the point through an absurdly simplified example. A stock trades at \(S = 100\) today and, in one year, will be worth either \(S_u = 120\) or \(S_d = 80\). Only these two outcomes are possible and we don't know with what probability each one will occur. In addition, the risk-free rate is \(r = 5\%\). It represents the return that a riskless asset can obtain. Finally, there's a call option with strike \(K=100\) on the stock. The question: how much is it worth?

DFA
Fig. 1. A simple setting of a stock and a call option. The stock can go up or down with probability $p$ and $1-p$ respectively. The payoffs of the call option in both states are known.

The nature of \(p\). In reality \(p\) is the probability distribution over the stock's future. No one knows it. We neither know \(p\), nor all the factors that might determine it. We're dealing with unknown unknowns. For all purposes, the future is epistemically uncertain. To get a sense of where it comes from we can try to trace the dependencies in reverse. The price of a stock changes whenever somebody decides to buy at the last market price, or whenever people readjust their limit orders and a stock exchange opens the floor for trading. One step backward, people adjust their orders because they update their expectations about whether it would be profitable to do so. One step back this happens for example because:

  • Global macroeconomic factors shift, or industrial shocks, scientific breakthroughs, regional wars or any political event happens. This is already hard enough to predict.
  • Expectations about future earnings or the future price change. This is where bubbles, feedback loops, fads, and public manias live.
  • People's risk-preferences change. A person who is very risk-averse could be unwilling to hold an asset, even if it's clear that its mean return is much greater than the risk-free return.

Every investor has an individual demand for the stock, which is based on his unique risk tolerance and willingness to endure uncertainty. From the aggregation of all individual demands one obtains the market demand for the stock, which also aggregates the bidders' individual risk-preferences. Hence, if people become more risk-averse, the price drops and the probability of the stock price increasing to a given level decreases (or so it is convenient to believe). The point is that the probability \(p\) is the real-world probability, which aggregates incredibly complicated causal dependencies based on expectations and individuals' risk-profiles. It is hubris to believe we know it.

Hence, in our simple setup the combination of \(p\) and the future values \(S_u=120\) and \(S_d=80\) implicitly contain information about the static real-life earnings expectations, risk-preferences, economic outlooks and so on.

Arbitrage. If we can't know \(p\), how do we determine the worth of the call option? Luckily, there is a way, by holding the right amount of cash (or equivalently a risk-free bond) and the underlying stock, such that this portfolio perfectly replicates the payoff of the call option. From the standard option formula \(\max{S-K, 0}\), we know that in the up-state it is worth \(C_u = 20\) and in the down-state \(C_d = 0\). Now, let's consider that we hold \(b\) units of a riskless cash bond and \(s\) units of the stock. We want to find the value \((s, b)\) that obtains the same payoff.

$$ \begin{bmatrix} S_u & 1+r \\ S_d & 1+r \\ \end{bmatrix} \begin{bmatrix} s \\ b \end{bmatrix} = \begin{bmatrix} C_u \\ C_d \end{bmatrix} \ \ \Rightarrow \ \ \begin{cases} 120s + 1.05b = 20 \\ 80s + 1.05b = 0 \end{cases} \ \ \Rightarrow \ \ \begin{bmatrix} s \\ b \end{bmatrix} \approx \begin{bmatrix} 0.5 \\ -38.095 \end{bmatrix} $$

With this exact combination of holding half a unit of stock and borrowing \(38.095\), in the good state we get exactly \(20\), and in the bad state exactly \(0\), replicating the payoff of the call option. What is the cost of entering this replication strategy? Borrow \(38.095\) and use \(11.905\) of your own money to buy half a unit of the stock. At the end it's worth either \(60\) or \(40\), and after repaying \(40\), you end up with either \(20\) or \(0\). The cost of this is \(50 - 38.095 = 11.905\). Call this number \(V\).

Consider what happens if the call option is cheaper. If \(C < V\), you'd want to buy the call and short the cash-stock combination. So you first borrow half a share, sell it for \(50\) and lend out \(38.095\). This leaves you with \(V=11.905\). With these buy the call option. At the end of the period, you receive \(40\) from the loan and \(20\) or \(0\) from the call option. This is enough to buy back the stock and return it to its original owner to complete the short. You just pocketed \(V - C\) without incurring any risk.

A similar arbitrage argument holds when \(C > V\). The point is that one needs \(C=V\) in order to avoid eternal riskless profit. So that's our valuation mechanism: the value of the call has to be the cost of replication in terms of the underlying stock and a risk-free asset.

Implications. No-free-lunch describes the peculiar nature of the world we live in. A contradiction between prices is also an invitation. In that sense, market efficiency is a restless, decentralized process of error-correction (see also Fama's efficient-markets formulation). Its order depends on people willing and able to search for discrepancies, bear execution and funding risk, and enforce contracts. And this process is never perfect. If it was perfect, and prices already contained all available knowledge, nobody would pay to discover or verify facts in the first place. Hedge funds would not be profitable, whereas in the real world they are, presicely because they have a social utility in harvesting the arbitrage opportunities and keeping markets close to perfect efficiency. So a perfectly efficient market is considered impossible. Errors may exist, but the errors large and accessible enough to finance their own correction tend not to survive.

There is broader similarity to other systems here. The closest physical analogue is thermodynamic equilibrium, especially when expressed in terms of free-energy gradients or chemical potentials, and detailed balance. A difference in an intensive quantity creates an opportunity for a flow; the flow exploits that difference and, in doing so, tends to destroy it. Of course, systems can remain in long-term disequilibrium when outside forces continually supply energy or impose constraints. A good example are states, which can distort prices and lead the system into unnatural behavior.

Risk-neutral pricing. Consider that in all the valuation calculations we didn't use \(p\). The arbitrage logic doesn't need to know the real distribution \(p\)! Now let's see another trick. We'll find another probability \(q\), such that the expected stock payoff is equal to the risk-free return. We solve

$$ q S_u + (1-q)S_d = S(1+r) \ \ \Rightarrow \ \ q = \frac{S(1+r) - S_d}{S_u - S_d} = 0.625. $$

Hence, if the probability of the up-state were \(0.625\), then the stock would be on average \(\mathbb{E}_q[S] = qS_u + (1-q)S_d = 105 = S(1+r)\) and the call option would pay out \(\mathbb{E}_q[C] = qC_u + (1-q)C_d = 12.5\). The discounted price is then \(C = 12.5/1.05 = 11.905\), the exact value we got from the no-arbitrage argument! This is nice because we obtained the same valuation but again without knowing the real probability \(p\).

Let's unpack. \(q\) is chosen so the stock earns the risk-free rate in expectation. Then, for the option, discounting its expected payoff under \(q\) yields the same price enforced by no-arbitrage. The distribution \((\text{up} \mapsto q, \ \ \text{down} \mapsto 1-q)\) is called risk-neutral because, under it, investors behave as though they require no extra expected return for bearing risk. Note that real investors are not indifferent to risk. And \(q\) does not forecast what will happen! It is only a set-up of mathematical probabilities at which current prices are consistent. This means that no one can earn above the risk-free rate on average.

Incompleteness. Let's consider also the case with three states: up \(S_u = 120\), middle \(S_m = 100\) and down \(S_d = 80\). As before, we have \(S=100\), risk-free rate \(r=5%\), strike price \(K=100\), which implies \(C_u = 20\), \(C_m = 0\) and \(C_d = 0\). In this case the risk-neutral probability measure is a triple \((q_u, \ \ q_m, \ \ q_d)\). Upon trying to solve

$$ \begin{cases} 120 q_u + 100 q_m + 80 q_d = 105 \\ q_u + q_m + q_d = 1 \end{cases} $$

one obtains infinitely many risk-neutral measures, parameterized for example as \((0.625-q_m/2, \ \ q_m, \ \ 0.375 - q_m/2)\) with the domain \(q_m \in [0, 0.75]\). As a consequence, the call option price can range anywhere from \(4.76\) to \(11.905\). This quickly shows that in an incomplete market (one in which there are fewer tradeable assets than number of possible future states), the no-arbitrage condition no longer selects a unique price for every payable claim: it only identifies a range consistent with the traded assets, while it is preferences, supply, and demand that determine where within that range the option trades.

Stochastic discount factors. We have so far deliberately avoided the real-world probability \(p\). To see where risk preferences enter, imagine an investor who consumes \(c_0\) today and will consume either \(c_u\) or \(c_d\) next year. Their expected utility \(\mathbb{E}_p[u(c)]\) is

$$ \mathbb{E}_p[u(c)] = u(c_0) + \beta\big[p\,u(c_u) + (1-p)\,u(c_d)\big], $$

where \(u\) captures the diminishing usefulness of additional consumption and \(\beta\) captures impatience. Now offer this investor a tiny amount \(\varepsilon\) of a claim \(X\): it costs \(\varepsilon X_0\) today and pays \(\varepsilon X_u\) or \(\varepsilon X_d\) next year. At a price they are willing to trade, the gain and loss in utility must balance at the margin:

$$ \begin{aligned} \mathbb{E}_p[u(c + \epsilon X)] &= u(c_0 - \epsilon X_0) + \beta\big[p\,u(c_u + \epsilon X_u) + (1 - p)\,u(c_d + \epsilon X_d)\big], \\ \Rightarrow \ \ \frac{\partial \mathbb{E}_p[c + \epsilon X]}{\partial \epsilon} &= -X_0 u'(c_0 - \epsilon X_0) + \beta \big[p\,u'(c_u + \epsilon X_u)X_u + (1-p)\,u(c_d + \epsilon X_d)X_d \big] \\ \Rightarrow \ \ 0 &= -u'(c_0)X_0 + \beta\big[p\,u'(c_u)X_u + (1-p)\,u'(c_d)X_d\big] \text{ as } \epsilon \rightarrow 0. \end{aligned} $$

Rearranging gives the pricing equation

$$ X_0 = \mathbb{E}_p[mX], \qquad m_s = \beta\frac{u'(c_s)}{u'(c_0)}, \quad s\in\{u,d\}. $$

The random variable \(m\) is called a stochastic discount factor. It is the rate at which the investor exchanges consumption in each future state for consumption today. It is high in a state where future consumption is scarce and an extra dollar is especially useful. In a market with many investors, the logic is the same: prices reflect the marginal value of money in each state. So we see that valuation depends on both the real probability \(p\) and the subjective psychological preferences \(m_s\) in each possible future state \(s\).

This also reveals the connection to the previous paragraphs. The bond, which pays \(1.05\) in either state, satisfies \(1 = \mathbb{E}_p[1.05m] = 1.05p\,m_u + 1.05 (1-p)\, m_d\). We can recognize a probability distribution \(q\) where \(q_s = 1.05p_s m_s\). They sum to one precisely because the bond is priced correctly. Substituting them back into the SDF equation gives

$$ X_0 = \frac{1}{1.05}\big[qX_u + (1-q)X_d\big] = \frac{1}{1.05}\mathbb{E}_q[X]. $$

So \(q\) is the real-world distribution \(p\), tilted toward states in which money matters more. This is why \(q\) is not a forecast: it is what remains after risk preferences have been folded into the probabilities. Claims that pay in bad states are valuable insurance and therefore expensive; claims that pay chiefly in good states must offer a higher average return under \(p\) to persuade people to hold them. Replication reaches the same prices without ever estimating \(p\), \(u\), or \(m\) because it exploits only the price consistency that those underlying forces produce.

Conclusion. What should a claim with an uncertain future payoff cost today? In an arbitrage-free market, its price is constrained by the prices of other traded payoffs. Risk preferences still matter, because a dollar received in a bad state is worth more than a dollar received in a good one; the stochastic discount factor records that difference. Risk-neutral probabilities package those state-dependent values into a convenient pricing rule, rather than describing anyone's literal forecast of the future.